Hey,

Your strategy probably has a rule that keeps you out of bad markets.

• Buy only when the index is above its 200-day moving average.

• Or only trade when the trend points up.

Traders call this rule a regime filter, and it is usually applied the same way everywhere.

You run your backtest and your equity curve looks better with the rule than without it, and the rule stays in the strategy forever.

I ranked 20 of these rules properly across

• SPY

• QQQ

• Bitcoin

Through four bear markets, over 2,700 backtests.

Good news: regime filtering works.

Some popular filters charge 4% a year for protection you could get cheaper, one spends five days out of six in cash, and several turn out to protect you from only one kind of disaster.

Your rule is somewhere on this list. By the end of this article you will know its rank, its price, and whether something on this page does its job better.

Why the Regime filter matters so much

If you already trade systematically, skip ahead. If not, here is why this one rule carries so much weight.

A regime filter is the switch that decides whether your strategy is allowed to trade at all. A dip-buying strategy, for example, buys pullbacks, and in a bull market pullbacks bounce.

The filter has one job. It tells those two worlds apart and shuts the system off in the wrong one. But the question is never whether a filter does this perfectly.

The question is what it protects you from, and what it charges you for that.

How the ranking works

Picture buy-and-hold SPY since the year 2000.

It paid about 7% a year, and at the worst moment a $100,000 account was showing $50,000.

That is the deal every filter is trying to improve, keep the 7%, skip the half-your-account part.

So every filter gets graded on four plain questions:

1. How good is the deal? Return compared to the pain. Skipping half the crash while keeping the return doubles this score.

2. How much of each crash did it dodge? Measured separately in the four big bears: the slow dot-com grind, the 2008 crisis, the three-week COVID crash, and 2022.

3. What does it charge? How much yearly return you give up versus just holding. This is the insurance premium.

4. Does it travel? A real pattern should work on QQQ and on Bitcoin too, not only on the S&P. A rule that works on one chart did not learn the market. It memorized that chart.

And one rule I set before running anything: no filter gets judged by its best setting. Every filter ran across its whole range of sensible settings, SMA lengths from 100 to 300 and so on, and gets scored on the middle result of all of them.

I proved in a separate 3,000-backtest study that the “best” setting from a backtest does not repeat in the future.

So the scores below are what you can expect with any reasonable setting, not what a lucky number once produced. (Settings under about 100 days were excluded on purpose: price dropping below its 20-day average is a dip you buy, not a regime you flee.)

The test itself is simple and the filter is the strategy:

• fully in the index while the rule says yes,

• fully in cash while it says no, trades at the next open,

• commissions included.

The full scorecard of all 20 filters sits in one comparison table near the end, so you can re-rank them by whichever score matters most to you.

Here is the ranking, worst to best.

The overpriced tier: #20 to #16

These five charge real money, 2 to 4 points of yearly return, and deliver protection you can get cheaper higher up the list. Three of them also stop working outside the S&P.

#20 - Price above TEMA

Parameters:
	maLen:	from 100 to 300 step 20

Data:
	Regime:	C > TEMA(C, maLen)

The triple-smoothed moving average, the fanciest formula in the test, and the most expensive failure.

#19 - Price more than x% above the n-day low

Only be long when price has already rallied x% off its long-term low. The crash numbers look fine until you see how it gets them, it spends one day in six in the market. That is not a filter, that is barely owning stocks, and it still costs 4 points a year. Of the three markets, it earned its keep only on Bitcoin.

Parameters:
	lowLen:	from 126 to 252 step 126
	pctLow:	from 10 to 40 step 5

Data:
	Regime:	C > (1 + pctLow / 100) * LLV(C, lowLen)

#18 - A fresh n-day high within k days

Has the market made a new high recently?

Against slow bears this is genuine armor.

It avoided 79% of the dot-com bear and 85% of 2008, close to the best in the test. Then COVID arrived, falling straight from an all-time high, and this filter caught 7% of it. Seven. Add a 3-point yearly premium for sitting in cash half the time, and you have a specialist, superb against the disaster that takes months, blind to the one that takes three weeks.

Parameters:
	hhLen:	from 63 to 252 step 63
	freshLim:	from 5 to 40 step 5

Data:
	Regime:	SinceHigh(C, hhLen) < freshLim

#17 - Price above Hull MA

The Hull average is built to hug price with almost no lag. Like TEMA, it is a flash-crash specialist, 86% of COVID dodged. And like TEMA, the whipsaw bill eats the rest, on SPY it ends up with half of buy-and-hold’s deal while charging 4 points a year.

Parameters:
	maLen:	from 100 to 300 step 20

Data:
	Regime:	C > HMA(C, maLen)

#16 - Higher lows: rising n-day low

Be long while the recent lows keep rising. Sounds like the textbook definition of an uptrend, and on SPY it is passable. Then you change the chart. On QQQ it adds exactly nothing, and on Bitcoin it makes things worse. The clearest one-chart artifact in the test.

Parameters:
	nLen:	from 20 to 60 step 5

Data:
	Regime:	LLV(L, nLen) > LLV(L, nLen)[nLen]

The middle tier: #15 to #11

#15 - Price within x% of its n-day high

Stay long while price holds near its recent highs. Cheap, works on all three markets, and barely protects: it sat through 73% of the dot-com bear and 90% of 2022, because with a typical 81% of days in the market it is the filter that barely filters. You pay little and you get little.

Parameters:
	hhLen:	from 126 to 252 step 63
	pctHigh:	from 80 to 96 step 2

Data:
	Regime:	C > pctHigh / 100 * HHV(C, hhLen)

#14 - Trend quality: how straight is the line

Be long only when price rises in a clean, orderly line.

Parameters:
	cqLen:	from 63 to 252 step 63
	cqThresh:	from 30 to 80 step 10

Data:
	Regime:	Correl(C, Days(20000101, BarDate), cqLen) > cqThresh / 100

#13 - Price above KAMA

Kaufman’s adaptive average speeds up in trends and slows down in chop.

The design genuinely works: it avoided 86% of COVID, the disaster every fixed-speed filter slept through. But adaptivity has a bill too, and over the full 25 years on SPY its deal lands below just holding.

Parameters:
	erLen:	from 10 to 50 step 10
	slowLen:	from 20 to 40 step 10

Data:
	Regime:	C > KAMA(C, erLen, 2, slowLen)

#12 - Price above SMA, plus or minus a band

A plain SMA filter with a tolerance band to cut whipsaws. Decent numbers everywhere, and one detail that should bother you, of this whole tier, its results swing the most when you change the settings. The band is a second dial, and the second dial added wobble, not safety.

Parameters:
	maLen:	from 100 to 300 step 50
	bandPct:	from -3 to 3 step 1

Data:
	Regime:	C > Avg(C, maLen) * (1 + bandPct / 100)

#11 - Price above WMA

A weighted moving average, recent days count more. Solid protection, 84% of 2008 dodged, works on all three markets, charges 2 points.

Parameters:
	maLen:	from 100 to 300 step 20

Data:
	Regime:	C > WMA(C, maLen)

Eleven filters down, and a quick recalibration before the top ten.

If your filter has not appeared yet,

it made the top ten, and right now you do not know what it is charging you or which rule beats it.

Below the line:

• The full top 10

• The complete 20-filter scorecard in one table

• And the one I would actually run.

The 2,700 backtests are done. Your part is copy-paste.

Read more

Hey!

There is a certain kind of AI trading answer that looks useful for about 15 seconds.

Ask ChatGPT or Claude, “Give me a profitable mean-reversion strategy,” and you usually get the same recipe with different wording: buy weakness, sell the bounce, add RSI, maybe add a volatility filter, optimize the lookback.

Nothing is technically wrong with that answer, which is why it wastes time.

It sounds good, but there is no research edge in it.

You are getting cleaned-up trading rules that has already been copied across forums, blogs, and prompt examples for years.

For that part of the workflow, I use Google NotebookLM.

Why NotebookLM is Different?

I do not use NotebookLM like a smarter ChatGPT.

I use it like a research notebook that can talk back, with citations.

The important difference is that NotebookLM works from the sources I upload.

I can give it white papers, RealTest documentation (RealTest is a backtesting software), old notes, trade journals, strategy writeups, and research PDFs. Then the conversation starts inside my material instead of inside the open internet.

That matters in trading, because the open internet is full of recycled strategy advice. If the sources are weak, NotebookLM will still produce weak output. But if the sources are good, it becomes a way to interrogate your own library faster.

ChatGPT, Claude and Gemini have access to the entire internet, including the latest news. That sounds like an advantage. In practice, it means they have more material to hallucinate from. The more data a model can pull from, the harder it is to know where an answer actually came from. NotebookLM cannot do that. It is locked to your sources.

And the basic version is free, so there is no reason to skip it.

The Multi-Paper Synthesis Method

Most of the interesting ideas do not come from one paper.

They come from putting several papers next to each other and asking where the logic overlaps.

I call this Alpha Cross-Pollination.

I might upload papers on Breaking the Trend, Momentum Factor, optimal rebalancing, stock mispricing, and volatility targeting into the same NotebookLM. These topics are related, but they are not usually presented as one strategy framework.

Then I usually ask questions that force NotebookLM to compare mechanics:

• Where do these papers point to the same market regime?

• Are there shared time windows, holding periods, or rebalance rules?

• Which filters reinforce each other?

• Which filters conflict?

• Is there a price-action signature that appears across more than one source?

• Which parts are directly testable in RealTest?

This is the part I care about because NotebookLM can scan 500 pages of technical material in seconds and pull out the overlapping pieces.

I am not asking it to create alpha/holy-grail out of thin air. I am asking it to help me notice where different pieces of research are quietly pointing at the same testable mechanic.

That beats asking an AI model to invent a strategy from scratch.

The Underrated Feature: Audio Overview

One of the most underrated NotebookLM features is Audio Overview, an AI-generated podcast based on your uploaded sources.

This is perfect for learning complex trading ideas away from the screen.

Before a walk, workout, or drive, generate the Audio Overview and listen to the strategy like a private quant podcast.

Quantopian does something similar with NoteBookLM audio. Explainer-style strategy content. (Example below)

What Else NotebookLM Can Generate

Audio is only one format.

NotebookLM can also turn the same sources into mindmaps, reports, slide decks, infographics, flashcards, quizzes, video overviews, and other study-style outputs.

For important papers, I like having more than one way to study the same source. A single PDF can become a written summary, a visual explainer, a short quiz, and an audio briefing.

I do not use every format every time, but the variety helps when the material is dense and complex.

Traders who use it can really outperform those who don’t.

Below I show you:

• How I use NotebookLM inside my AI research workflow instead of using it as a random AI chatbox.

Read more

TQQQ is easy to sell with one screenshot.

Start with $10,000. End with millions. Make regular index investing look slow.

Then leave out the years where the account gets cut by more than 80%. Leave out the trade list. Leave out the moment a real trader starts reducing size, overriding signals, or deciding the backtest was never realistic.

My test is smaller and more practical:

• define the rule,

• trade only after the signal is known,

• include costs,

• compare against buy and hold,

and ask whether the drawdown is still something a human can follow.

A leveraged ETF multiplies the daily move of an index. UPRO targets 3x the daily S&P 500 move. TQQQ targets 3x the daily Nasdaq-100 move. The daily reset is the catch. Smooth trends help, choppy markets and bear markets punish the path.

I tested one practical question in RealTest: can a moving average reduce leveraged ETF risk enough to make the trade-off worthwhile?

From 10,000 to over 2 million. The return chart is why traders care.

But… the drawdown is why the strategy needs rules.

The paper won the 2016 Charles H. Dow Award

Michael Gayed and Charlie Bilello made that argument in Leverage for the Long Run: A Systematic Approach to Managing Risk and Magnifying Returns in Stocks.

The credibility matters. The paper won the 2016 Charles H. Dow Award from the CMT Association.

The idea is simple: volatility is the enemy of leverage. Moving averages may help identify when the market environment is friendly enough to use leverage.

I tested that in RealTest.

I tested the idea with SPY, QQQ, SSO, UPRO, and TQQQ

The test uses:

• SPY for the S&P 500

• QQQ for the Nasdaq-100

• SSO for 2x S&P 500 exposure

• UPRO for 3x S&P 500 exposure

• TQQQ for 3x Nasdaq-100 exposure

The regime signal comes from the normal ETF. SPY controls SPY, SSO, and UPRO. QQQ controls QQQ and TQQQ.

The rule is simple: hold the ETF when the signal ETF is above its 200-day moving average. Exit to cash when it is below.

The signal is known at the close. The trade happens at the next open. I also included commissions and slippage. That matters because same-close execution and zero-cost assumptions can make a strategy look cleaner than it is.

Each strategy and benchmark starts as its own $10,000 sleeve and compounds separately. Treat the results as side-by-side research, not as one live portfolio trading every variant together.

The moving average helped SPY and QQQ

The unleveraged results were good.

SPY buy and hold produced a 6.92% CAGR with a -48.56% max drawdown. The moving average version produced a 7.74% CAGR with a -22.62% max drawdown.

QQQ buy and hold produced an 8.19% CAGR with a -80.14% max drawdown. The moving average version produced a 10.40% CAGR with a -27.02% max drawdown.

On SPY and QQQ, the filter did what a trend filter is supposed to do. It reduced drawdowns dramatically. For a trader trying to follow a system live, that matters.

But leveraged ETFs change the trade-off.

The moving average made TQQQ safer, but also much poorer

Buy and hold TQQQ finished at about $3.96 million from a $10,000 sleeve. The max drawdown was -81.61%.

The moving average version finished at about $1.12 million. The max drawdown was -57.19%.

The moving average made TQQQ less dangerous and gave up most of the ending wealth.

The takeaway: lower drawdown still has to pay for itself.

A slow trend filter can sell after a large decline has already happened, then buy back after the strongest part of the recovery has already happened. With a normal ETF, that is annoying. With a 3x ETF, it can be extremely expensive.

Whipsaw is the reason here.

The Working Layer

Free takeaway: a moving average can make leveraged ETFs less dangerous. The safer version still has to earn its place.

Below is the part you can actually use:

• Exact rules

• Full results

• Checklist

• Decision model

• RealTest script

Subscribe now if you want the tested version, not just the idea.

Read more

If you have tried using AI for trading research, you know the trap.

The first AI answer can look amazing.

Then you spend the next hour:

• fixing syntax

• waiting the AI reponse

• checking the rule logic

• struggling with AI hallucination

• pointing the agent back to the right data

• rerunning the backtest

• etc.

So I gave Codex, Gemini, and Claude Code the same job:

Create the strategy, run the base backtest, optimize the moving-average settings, run a walk-forward test, and produce charts I could review.

Which one wins? Let’s find out.

Same task, same data, same budget

I used the normal $20 plan for each tool. Agents were:

• OpenAI Codex, using GPT-5.5

• Gemini 3.1 Pro

• Claude Code, using Opus 4.7

Each agent worked in the same local RealTest environment and received the same trading idea.

The strategy was a very simple:

• Trade SPY only.

• Buy when the 20-day moving average crosses above the 50-day moving average.

• Only buy when SPY is above the 200-day moving average.

• Exit when the 50-day moving average crosses back above the 20-day moving average.

I did not choose this because it is a special edge. I chose it because a simple strategy makes the comparison cleaner.

What each AI had to complete

The first job was the base backtest. The agent had to create a RealTest script, use the local SPY data, run the test, and produce the first equity curve.

The second job was optimization. I did not want the agent to assume 20, 50, and 200 were automatically the best settings. I wanted a grid of moving-average combinations and a heatmap showing the shape of the result.

Because a single best MAR number can fool you. If good results only appear in one tiny corner, the strategy may be fragile. If nearby settings also work, the idea is easier to keep researching.

The third job was walk-forward testing. The agent had to optimize on one historical window, then apply the selected parameters to later data.

The exact prompts I gave them

Promp 1:

Write a RealTest strategy for SPY only. 
Buy SPY when the 20-day moving average crosses above the 50-day moving average, but only when the close is already above the 200-day moving average. 
Go all-in with 100% of capital, this is a single-position strategy, always either fully in or fully out. 
Exit when the 50-day moving average crosses back above the 20-day moving average (death cross). 
Long side only. 
Test from 2000 to 2026.
Use DataFile:    spy.rtd to import SPY symbol

Prompt 2:

Now add an optimization to that strategy. Optimize these three parameters:
Fast MA period: test values 10, 15, 20, 25, 30
Slow MA period: test values 40, 50, 60, 75, 100
Trend filter MA period: test values 150, 200, 250
Skip any combination where the fast MA period is greater than or equal to the slow MA period — those are invalid. Run the optimization from the command line without any dialogs. 
Show the results in a 2d matrix heatmap plots (use MAR).

Prompt 3:

Now set up a walk-forward optimization on that same strategy. 
Use a 3-year in-sample window, stepping forward 1 year at a time, measured in years. 
Keep only the single best parameter combination per window. 
Run it in two steps from the command line: first the optimizer pass, then a separate backtest pass that uses the walk-forward results. 
The backtest results should show both in-sample and out-of-sample rows so we can compare them.
Show the results in an ISS + OOS plot. 

Codex

First, I gave my first prompt to Codex.

Codex started by inspecting the local project and checking the data source before writing the strategy file.

It imported SPY from Norgate data, ran the base backtest, and produced an equity curve.

I still opened the result in my own RealTest setup. The agent can write and run the script, but the trader has to confirm the test matches the intended rules.

For the optimizer, Codex built the moving-average grid, ran the command-line optimization, and generated a heatmap.

It also created the walk-forward version and produced a chart comparing in-sample and out-of-sample equity curves.

Overall, Codex did a very good job and completed every task.

Gemini

Next, I wanted to test Gemini. I had heard that some RealTest users also use Gemini models. How good is it compared to others? Let’s find out.

Gemini worked through the same RealTest environment inside Antigravity.

It inspected the local files, generated the strategy script, ran the base backtest, and produced the first equity curve.

The useful detail in Gemini’s optimization pass was that it respected the moving-average constraint: the fast average has to be shorter than the slow average.

That sounds obvious, but it matters. An optimizer chart is not useful if the grid includes invalid strategy logic.

Gemini tested fast averages, slow averages, and the long-term trend filter, then produced the heatmap.

Gemini also ran the walk-forward test. The out-of-sample curve was lower than the in-sample curve.

Subscribe now

Claude Code

Next, the test is Claude Code. Claude is probably the most well-known AI coding agent.

Claude Code handled the same sequence with the least hand-holding.

It read the files, imported the data, ran the base backtest, and generated the first equity curve. I checked the result inside RealTest, and the base test worked.

Then it used my existing optimization flow, ran the parameter grid, and created a heatmap comparable to the other agents.

The walk-forward step was the hardest part of the comparison because it was not just “write a script.”

Claude Code had to write rolling parameter selections into the script, run the optimizer pass, and then run a separate backtest pass using those selected parameters.

It produced the final chart showing the difference between optimized in-sample performance and out-of-sample performance.

All three worked. One was better?

All three agents worked.

This was not a demo where two tools failed and one tool looked heroic.

Codex did the base backtest, optimization, and walk-forward test.

Gemini did the base backtest, optimization, and walk-forward test.

Claude Code did the base backtest, optimization, and walk-forward test.

The difference was how much time it takes.

In this RealTest setup, Claude Code felt faster, steadier, and easier to keep inside the research process. It needed fewer reminders and fewer manual pushes (because I’m not allowing AI to do everything on my computer).

Claude Code was my winner for this test.

And second and third place are tied between Gemini and Codex.

I also checked how the models I used rank across different performance benchmarks. Claude Code comes out on top, followed by Codex in second place and Gemini in third. You can check the rankings here if you’re interested.

My conclusion is this:

For this RealTest research task, Claude Code got me from idea to inspectable output with the least hand-holding.

Do not treat this as a permanent ranking

This was one strategy, one environment, and one workflow.

Model versions change. Tooling changes. Local file structure, permissions, data paths, and personal research habits all change the result.

So I would not treat this as a permanent ranking of AI tools.

I would treat it as a practical example of how to judge them.

Systematic traders should not look for a brand to believe in. We should look for a process we can inspect.

The real problem is starting from zero

The most annoying part of using AI for RealTest is starting from zero every session.

The agent does not know your folder structure. It does not know how you import data. It does not know which RealTest syntax mistakes it made last time. It does not know how you prefer to run backtests, optimizations, charts, and reports.

So the first part of the session becomes setup work.

You explain the same things. You correct the same assumptions. You remind it what RealTest expects. You push it back toward the research process you actually use. You waste so much tokens. And tokens = money.

That is why I built the new SetupAlpha course around Claude Code and RealTest.

The course is not based on the idea that AI makes 100% win-rate strategies. It is built around a narrower, more useful idea:

If the files, prompts, workflow, and RealTest structure are set up properly, Claude Code can help you plan a strategy, write the script, fix errors, run the backtest, create charts, and organize the research result.

That saves time. A lot of time.

More importantly, it helps more ideas reach the point where they can be judged.

You can start a research task, let the agent work through the mechanical parts, then come back and inspect what was tested, what failed, what passed, and whether the strategy deserves another round.

That is the right role for AI in systematic trading research.

Not replacing judgment.

Getting more ideas to evidence.

If you have tried using AI with RealTest before and felt like it wasted more time than it saved, this workflow was built for that exact problem.

Check out the course here

Bottom line

AI can write a trading strategy.

That is no longer the interesting question.

But the better question is whether it can help complete a research process that a serious/professional trader would respect.

I’m sure you want a clean reversal setup.

The one nice setup is where SPY gaps down, shakes everyone out, and then rips higher while the trade looks obvious in hindsight.

This idea is actually very simple.

• But how does it work?

• What are the rules?

• And should you even use it?

In this article, we’ll find out.

Gap Down Only Proves Weakness, Not Exhaustion

A gap below yesterday’s low tells you only one thing before the trade:

SPY is opening in a weak location.

It does not tell you if the sellers are exhausted or the move down is just getting started.

Every finished chart becomes too easy to explain.

1. Winners get called reversal setups.

2. Losers get called obvious warning signs.

That is not research.

So this test needs to answer the question before the outcome is known:

When SPY opens below the prior day’s low, is buying that weak open a repeatable trade, or is it just a pattern that looks better after you already know the close?

Below is another candle-related stock market pattern I researched before:

Yesterday I made a YouTube video where Codex, Gemini, and Claude Code all got the same RealTest task: write the strategy, run the backtest, optimize it, walk-forward it, and produce the charts. I wanted to see which agent could be the best.

The results were pretty interesting.

The Test: Buy the Weak Open, Hold Until Green

The hypothesis is that a gap below the prior day’s low can create short-term reversal pressure.

SPY opens below the entire previous day’s range. Overnight or pre-market selling has pushed it into a weak location. If that selling is forced, emotional, or simply overdone, the regular session may reverse part of the move.

The strategy is only trying to capture that reaction. Macro conditions and multi-month SPY exposure are outside the scope of this test.

The test asks one narrow question:

If SPY opens below yesterday’s low, is there enough short-term mean reversion to buy the open and hold until SPY prints a green day?

Narrow tests are useful because they leave less room for vague explanations. Either the behavior shows up in the data or it does not.

One Entry Rule. Zero Filters.

The strategy buys SPY after a weak open.

Buy SPY when the next regular-session open is below the previous regular-session low.

More about the rules (step-by-step) below.

There are no moving averages, RSI filters, volatility filters, or optimized gap-size thresholds. A one-condition entry gives the result fewer places to hide. If the test works, you know what drove it. If it fails, you do not have twelve filters to untangle.

But before we look at the results, one practical note.

If you have used AI for RealTest work, you probably know the annoying part:

• You use AI thinking it will make everything faster, but in the end you realize you wasted even more hours fixing errors and don’t understand any of the code AI wrote.

That is the workflow problem I built the new SetupAlpha course around.

The course connects Claude Code to a RealTest. My goal is not to make AI predict the market. My goal is to shorten the loop from trading idea to tested RealTest script.

Check out the course here.

It Didn’t Beat Buy-and-Hold, But Exposure Changed the Story

Here are the RealTest summary stats (includes slippage):

This strategy made far less money than buying and holding SPY.

SPY also had the higher Sharpe ratio. If the goal is to maximize long-only SPY exposure, this test is not a replacement.

The more useful comparison is exposure. The strategy earned a similar MAR while being invested only 21.46% of the time. That changes the interpretation and I would treat it as a short-term system, not as a “better SPY” strategy.

That kind of system can be useful as a portfolio component, because it does not require you to be invested every day.

Next let’s talk about:

• Trading Strategy Rules (Step-By-Step)

• How You Can Use It

• RealTest Strategy Script

• What Do We Learn

Read more

Hey Trader,

You’ve probably heard that before:

Trend following and mean reversion are usually framed as a contest.

• Which one wins?

• Which one has the better Sharpe?

• Which one survives 2026?

That’s the wrong test.

A trend system gets paid when movement persists. A mean-reversion system gets paid when movement snaps back. One wants continuation. The other wants exhaustion. If you only compare them as standalone systems, you miss the main reason both still matter.

I already covered the direct comparison in Trend Following vs. Mean Reversion: Which Strategy Wins in 2026?. But this test goes one layer deeper.

Opposite Edges

Trend following can look broken for a long time before it pays.

It can lose often, then make a large amount when a move extends. The hit rate can look bad. The equity curve can go quiet. The strategy survives because the winners are supposed to be much larger than the losers.

That is why trend following is so hard to hold. A system with a low win rate feels broken before the payoff arrives.

Mean reversion is the opposite shape.

It often wins more frequently. It gets paid when short-term moves stretch too far and reverse. The problem is not usually the number of losses. The problem is the loss that happens when price does not revert and a short-term dislocation turns into a real trend.

That is why mean reversion is good until happens a very negative event.

This is the structural difference:

• Trend following: lower win rate, larger right-tail winners, many small failures.

• Mean reversion: higher win rate, smaller normal wins, occasional left-tail damage.

Hamill, Rattray, and Van Hemert describe trend following as positively skewed, meaning the payoff shape can include many small losses and occasional larger winners. AQR’s long-horizon trend-following research also treats trend following as a persistent return style, not as a one-period anomaly.

On the mean-reversion side, Poterba and Summers showed that stock-return behavior changes by horizon. Short horizons and long horizons do not have the same serial structure. That matters because mean reversion is never a universal property. It is a horizon-specific bet.

So the question is not whether trend or mean reversion is “better.”

The question is whether the two return shapes fail differently enough to deserve capital in the same portfolio.

Last week, I published the Nasdaq momentum strategy.

These are the YTD results.

You can download RealTest script and strategy rules here.

All time high mean reverison YTD below:

The RealTest Result

I tested three versions:

Trend standalone: SPY trend-following logic.

Mean reversion standalone: S&P 500 stocks.

Combined shared 50/50: 50% trend-following exposure cap and 50% mean-reversion exposure cap in the same book.

Here is the RealTest-style stats output from the TF/MR test:

Quick definitions before the numbers:

• ROR is RealTest’s rate of return.

• Maximum drawdown (MaxDD) is the worst peak-to-trough loss.

• MAR is return divided by maximum drawdown. Higher means more return per unit of drawdown.

The combined version did not win by having the highest return. It won by changing the risk shape.

If you optimize only for standalone return, mean reversion looks better than trend following. If you optimize only for standalone drawdown, mean reversion still looks better. But the combined book had a lower drawdown than either standalone system and a higher MAR than both.

That is the portfolio doing its job.

The two edges did not need to predict better. They needed to lose at different times.

You can test our portfolio tool here: https://setupalpha.com/pages/realtest-strategy-combiner

Math

The math only works if the two strategies do not move together.

Correlation means how often two strategies move together. If two strategies have low correlation, their good and bad days overlap less.

If two strategies have similar volatility but less-than-perfect correlation, the combined volatility is lower than the simple average of the two standalone volatilities.

Correlation Is Not a Label

A strategy pair is not “uncorrelated” forever.

Correlation moves. It rises during certain regimes. It can rise exactly when you need it to stay low. That is why a single full-period correlation number is only the start.

The 252-day rolling correlation averaged 0.32. The latest value was 0.33. The range went from -0.07 to 0.67.

The full rolling chart shows the same thing visually:

That is a good result, but it is not a permanent guarantee. A 0.32 average correlation says the two systems usually moved differently enough to help each other. A 0.67 high says there were windows where the protection became weaker.

One mistake is to turn this into a slogan: combine two low-correlation strategies and move on.

The combined result can improve while the portfolio is still fragile. One leg can carry almost all the return. Losing months can line up during the same market stress. Correlation can look low during calm periods and rise during the exact window where you needed protection. The combined Sharpe can hide a tail that only shows up once every few years.

Before I would trust this kind of structure, I want four checks:

• Skew: did the combination actually reduce the worst payoff shape of each standalone system?

• Rebalancing: did monthly maintenance add anything, or was the 50/50 label cosmetic?

• Correlation: did the relationship stay low enough across rolling windows?

• Allocation: did the best split come from balance, or from secretly overweighting one edge?

If any one of those checks fails, the combined curve is less useful than it looks. A lower MaxDD is not enough by itself. I want to know whether the lower drawdown came from a real mismatch between the two systems, or from one lucky historical sequence where the bad periods happened not to overlap.

Next:

1. SPY Trend Strategy + SP500 Stocks mean reverion strategy

2. RealTest Script

3. How much of the result came from skew balance, rebalancing, correlation behavior, and allocation choice.

Read more

Has this happened to you before?

You added another volume filter.

• And your backtest improved.

You added a secondary RSI condition.

• The Sharpe ratio went up.

You kept going: a price filter, ATR expansion, a weekly momentum rank.

Each one made the backtest look better.

And then you took it live on IBKR.

The strategy didn’t behave the way the backtest said it would.

What a degree of freedom actually costs

Every rule in a trading strategy is a free parameter. A free parameter is a number or condition your system used to describe the past. The more parameters you have, the more precisely your model fits historical data. That precision is the problem.

Robert Pardo, whose book on strategy testing introduced the degrees-of-freedom concept to systematic trading, described it this way: a model with enough free parameters can fit any historical dataset, including one generated by random noise. The fit tells you nothing about what will happen next. It tells you how precisely you’ve described what already happened.

Each rule added to an entry condition is one more way the strategy is tied to the specific patterns that appeared in your training data. The strategy doesn’t get stronger. It gets more specific to a past that won’t repeat exactly.

The question is how large that cost is. I ran a test to find out.

The same strategy, two versions

I built a mean-reversion strategy on S&P 500 stocks.

Idea: stocks in an uptrend that pull back hard enough for RSI(2) to drop below 20, while the broader market (SPY) is also pulling back. Entry on a limit order at the Bollinger Band bottom. Exit when RSI(2) recovers above 70.

Four conditions. That’s the simple version.

Then I added eight more:

• Volume filter (average daily volume above 1 million)

• Price filter (above $10)

• Secondary RSI (RSI(5) below 40, alongside RSI(2) below 20)

• Bar pattern (previous close lower than the one before)

• Weekly momentum (price below where it was five sessions ago)

• Extension check (price not more than 15% above the 200-day average)

• ATR expansion (ATR rising versus five days prior)

• Secondary market filter (SPY RSI(5) also below 50)

Twelve conditions total. Each one justifiable on its own and each one adds a degree of freedom.

I tested both versions from January 2000 through December 2018 as in-sample, and January 2019 through April 2026 as out-of-sample.

What the in-sample data said

In-sample, the complex version looked better.

If you had stopped here, you would have deployed the 12-rule version. It had higher CAGR, smaller drawdown, stronger MAR and Sharpe, a higher win rate, and a higher profit factor.

The simple version still had better expectancy per trade. If you are optimizing a backtest, that one warning is easy to ignore.

Marcos López de Prado made the same observation about machine learning funds in his 2018 QuantCon talk.

ML algorithms, he argued, will always find a pattern if given enough parameters to search through, even when no real pattern exists. The result is a backtest that fits the training data with high precision and fails immediately afterward. The mechanism is identical whether you’re running a neural network or adding RSI conditions one by one.

Gary Antonacci showed the opposite end of this in Dual Momentum Investing. Two rules: relative momentum (which asset class is outperforming) and absolute momentum (is momentum positive at all). Tested across 40+ years of data covering multiple regime changes. The simplicity wasn’t a limitation but it was what allowed the edge to persist.

If you want to understand what a structurally simple system actually looks like before you start adding to it, this article covers two examples in detail: How Simplicity Turns $100K Into $1.7M with 2 Simple Trading Strategies. Understand the core structure before adding complexity.

The NASDAQ Momentum Rotation strategy I released last week is another example. One momentum + trend signal, one rotation rule.

You can see it here.

What happened out of sample

Out of sample, the results reversed completely.

The simple strategy improved out of sample. CAGR went from 7.18% to 11.78%. Sharpe went from 0.69 to 1.18. Win rate from 66.79% to 75.80%.

The complex strategy fell apart. CAGR dropped from 8.62% to 4.51%. Sharpe from 0.93 to 0.50. Expectancy per trade from 1.12% to 0.61%.

The eight extra conditions didn’t survive contact with a new market period. Each filter that improved the in-sample curve was fitting a pattern that existed in 2000-2018 data and didn’t carry forward to 2019-2026.

If you have a strategy with a lot of entry conditions, some of them are doing this.

Because the problem is you can’t tell which ones by looking at the backtest.

In-sample, they all look like genuine improvements.

The only way to separate structural rules from historical accidents is to ask a different question about each one.

• That’s what the paid section covers.

• RealTest strategy code included.

Read more

Hey Systematic Trader,

Does equity curve trading actually make your strategy better?

Is it just overfitting with extra steps, or something real that professional funds use every day?

And if it does work, why does every article tell you to stop trading when your equity drops, when the data says the opposite?

I had all these questions.

So I ran 25 years of backtest data on a mean reversion strategy, tested seven different parameter values, and compared the results side by side. This article is what I found.

• MaxDD dropped from 21.26% to 16.47%.

• Win rate climbed from 69% to 72%.

• Profit Factor jumped from 1.76 to 2.06.

Here is exactly why that works, and how to build it in RealTest.

What Is Equity Curve Trading?

Equity curve trading means using your strategy’s own recent performance as a filter for future trades.

The idea:

Watch your equity curve against a moving average. When the curve is trending up, the strategy is running hot, keep trading. When it is trending down, the strategy is cold, step aside.

The instinct is correct. The direction, for mean reversion, is completely wrong.

Why Mean Reversion Flips the Logic

Mean reversion is the rubber band trade. Pull a price too far from its average, it snaps back. You trade that snap.

The rubber band applies to the strategy’s own returns, not just to the prices it trades.

When a mean reversion system goes through a bad run, the edge is not broken. It is compressed. Think of a basketball player on a shooting slump. A cold week does not mean he forgot how to play. It means he is due. The law of averages says his next game looks better, not worse.

Mean reversion strategies work the same way. Losses cluster, then reverse. The bad run itself is the signal.

This is called negative return autocorrelation.

Long phrase, simple idea: the bad periods and good periods of a mean reversion strategy alternate. They do not trend in one direction.

So the filter runs backwards from what feels intuitive:

• Trade when equity is above its moving average: you enter after the warm streak is already winding down

• Trade when equity is below its moving average: you enter right as conditions are statistically turning

The Tracking Benchmark

Before the results, there is one piece of plumbing to understand.

The filter needs a tracking benchmark. Think of it as a shadow. Your strategy casts a shadow that keeps moving even when the strategy itself stands still.

Why do you need it? If the main strategy stops taking trades, its equity curve goes flat. A flat line has no meaningful moving average. You would be reading a signal from a frozen instrument. The shadow keeps moving, keeps recording what would have happened, and keeps the filter alive.

In RealTest, the setup is two strategy blocks:

Benchmark: ben
    EntrySetup:  Buy
    Using:       Common

Strategy: MeanReversion
    EntrySetup:  Buy and Extern(@ben, S.Equity < avg(S.Equity, 20))
    Using:       Common

@ben is the “shadow”.

On every potential entry bar, the strategy checks one thing: is the shadow’s equity currently below its 20-day average?

• Yes: take the trade. The strategy has been running cold. Statistically favorable territory.

• No: skip the trade. The warm streak is active. Let it exhaust itself first.

The Results (EQC Filter vs Nan)

Same strategy + period (Jan 2000 to Jan 2026). Same $100K starting capital.

Two metrics worth a quick definition.

• MAR is annual return divided by your worst drawdown.

• Profit Factor is total winning dollars divided by total losing dollars. A Profit Factor of 2.06 means for every $1 you lose, you win $2.06. Anything above 1.5 is solid.

The filter cut maximum drawdown by 22%. MAR improved from 0.44 to 0.49. Sharpe from 0.77 to 0.80. Win rate jumped 3 percentage points. Profit Factor went from 1.76 to 2.06.

Fewer trades (1,168 vs 1,974) and lower absolute net profit ($640K vs $912K). The strategy sits out a portion of the time.

If you care about smooth equity and risk-adjusted returns, the filter wins. If you care only about maximum raw compounding, the unfiltered version wins on net profit. Both are valid answers depending on what you are building.

Equity Curve: Side by Side

Here is what the tracking benchmark (green, always running) looks like overlaid with the filtered strategy (dark red):

They track closely for most of the period. The gaps are the moments the filter was active and the main strategy sat out.

The shadow running alone:

The filtered strategy alone:

Look at 2001 to 2002, and again at 2008 to 2009. The filtered curve holds a staircase shape during those periods. The strategy pauses, the worst drawdown days pass, then it resumes. The rough patches are shorter and shallower.

That visual tells you most of what you need to know. But there is still one question it cannot answer.

Is len=20 a Lucky Number?

If only len=20 works and everything else fails, this is overfitting. Not an edge.

So I ran the optimizer across seven values: 5, 10, 20, 50, 100, 150, 200.

Every single length produces a positive result. The equity curves cluster together. No lucky spike, no fragile optimum hiding in a narrow band.

Shorter lengths trade more and return more in absolute terms: longer lengths trade less and reduce drawdown further.

Both ends of the range behave sensibly.

The only value that starts to struggle is 200, and even there the strategy is profitable. The issue at len=200 is simply that 602 trades over 25 years is too few events for compounding to work.

The number 20 is not special. The mechanic underneath it is.

The Full Strategy Rules + RealTest Code

Below I am sharing the complete strategy rules and ready-to-use RealTest code with paid subscribers.

What you will get:

1. Complete RealTest code (copy-paste ready): full working script including the tracking benchmark

2. Exact entry and exit rules: every condition explained in plain English, step-by-step

3. All parameters explained: what each does and why it is there

4. Which len to choose and why: the MAR vs ROR trade-off broken down clearly

5. [MOST IMPORTANT] Why this filter backfires on trend following: and the costly mistake to avoid

Read more

Hey Trader!

• Are you start skipping signals on bad news days?

• You reduce size after a rough week?

You pause the system during a volatile stretch. Each decision feels reasonable in the moment. None of them are in the backtest.

Six months later the system is technically still running, but you’re not really trading it anymore. You’re managing, overriding and interpreting it. And the results look nothing like what you built.

This article maps out where that happens.

40+ specific points across every phase of running a systematic strategy, from building it to surviving a drawdown.

1. Research & Backtest Phase Mistakes

Before a single dollar is risked, the mistakes are already happening.

Mistake #1. You Added Complexity Because Simple Felt Like Giving Up

You had a working strategy. Clean signals, clear rules, low parameter count. Then you started “improving” it. A trend filter here. A volume confirmation there. A volatility regime switch. Now it has eleven parameters and none of them interact the way you think they do out-of-sample.

Think of it like a combination lock. One dial is easy to crack. Ten dials looks secure, but now ten things need to land exactly right before anything opens. Every parameter you add works the same way.

Simple strategies tested on 25 years of data beat complex strategies that look better in a single backtest. Every extra condition is another thing that has to be true simultaneously. More things means more ways to fail when the market does something new.

Mistake #2. You Got One Good Backtest. You Stopped Looking.

You run 50 parameter combinations. One produces a beautiful equity curve. You stop there.

The problem: that’s like memorizing the answer key instead of learning the subject. It aces the test it studied for. Every other test, it fails.

In RealTest, run a full optimization grid and look for flat parameter regions, clusters where adjacent values all produce similar results. A single peak surrounded by valleys is a warning sign, not a discovery. Flatness is the evidence of a real edge.

Mistake #3. Your Backtest Window Avoids Every Crisis That Would Have Hurt You

Traders unconsciously pick backtest windows that include conditions they remember.

“I’ll test from 2015.”

That excludes 2000 to 2002 and 2008, the two periods that would have shown the real risk profile. Always start from 2000 or earlier. If your system survives the dot-com collapse and 2008 without triggering your psychological exit threshold, you have real evidence. If you’ve never seen it in those environments, you have a theory.

Mistake #4. You Optimized for Maximum CAGR. You Built a Strategy You Won’t Survive.

CAGR is the most visible number so traders maximize it. But high CAGR almost always comes with a max drawdown that is psychologically impossible to live through. A strategy with 28% CAGR and -42% max drawdown will not be traded through a -42% drawdown. It will be stopped at -20%.

Design for survivability. The right target is Sharpe ratio or CAR/MDD (annual return divided by max drawdown). A system with 17% CAGR and -12% max drawdown will be traded for decades. A system with 28% CAGR and -42% max drawdown will be abandoned within eighteen months.

Mistake #5. You Backtested on Yahoo Finance. You Have Never Actually Tested Your Strategy.

Yahoo Finance only shows stocks that still exist. Every company that went bankrupt, was delisted, or merged away is missing. Your “S&P 500 strategy” never bought a company that went to zero, because those companies aren’t in the dataset. The results look clean because all the painful trades were silently removed before you ran the test.

Norgate Data includes delisted securities and historical index constituents. In RealTest, the InSPX filter uses the actual S&P 500 membership at each historical date, not the current index retroactively applied to the past. The difference between a Yahoo backtest and a Norgate backtest of the same strategy is often significant.

One is real. One is fiction.

Mistake #6. Your Strategy Works in Bull Markets. You Don’t Know What Else It Works In.

A strategy that performed well from 2010 to 2021 may have only worked in a sustained low-volatility bull market. That environment is not the default.

Test explicitly across:

• Dot-com collapse (2000 to 2002)

• 2008 financial crisis

• 2020 crash

• 2022 rising-rate bear market

• At least one sideways grinding period

If the strategy only produces acceptable results in some of these, you don’t have a strategy. You have a favorable-period artifact.

Mistake #7. You Treated Walk-Forward Analysis as Optional. It Isn’t.

Think of it like testing a recipe. You develop it in your own kitchen. Then you hand it to someone else, in a different kitchen, with ingredients they bought independently. If it still works, it’s a real recipe. If it only worked with your exact setup, it was a coincidence.

Walk-forward analysis does that for backtests. Optimize parameters on the first portion of history, lock them in, test on the next untouched portion, then roll forward. In RealTest, the Walk-Forward module runs this automatically.

If your strategy survives with 20 to 40% performance degradation out-of-sample, that’s genuine evidence. If it collapses entirely, the in-sample results were data mining, not discovery.

Mistake #8. You Modeled Transaction Costs as Zero (or as a Guess)

“I’ll just use $0.01 per share” is not a model. It’s optimism. A strategy showing 25% CAGR with no costs might show 14% after realistic costs.

In RealTest, model Interactive Brokers’ tiered pricing accurately:

Commission: Max(1.0, 0.005 * Shares)

Add LimitExtra to account for partial fill behavior. Run with realistic slippage. If the strategy doesn’t survive realistic costs, it doesn’t exist outside of the test environment.

2. Going Live Mistakes

The transition from backtest to live is where most strategies hit their first real obstacle.

Mistake #9. You Went Full Size on Day One

You backtested with $100,000. You go live with $100,000. The first three trades lose. The drawdown feels completely different than it did in the backtest. Real money creates an intervention urge the simulation never produced.

Go live at 25 to 50% of intended size for the first 30 to 60 trading days. You’ll experience real fills, real slippage, and the emotional texture of live losses before your full capital is committed. The data you collect in that period is the most valuable your system will ever produce.

Mistake #10. You Set Expectations from Paper Trading. They Were Wrong Before You Started.

Paper trading performance is almost always cleaner than live. No slippage, no partial fills, no psychological pressure. When live trading underperforms paper trading (which it always does), traders assume something is wrong with the system. Usually what’s wrong is the comparison.

Paper trading confirms execution logic. Live trading is where you learn what the strategy actually feels like to run.

Mistake #11. Your Live Orders Don’t Match Your Backtest Logic. You Haven’t Checked.

The most common going-live mistake is a silent execution error:

• Strategy says: buy at limit

• Broker received: market order

• Price came in 0.8% worse

• Nobody noticed because nobody ran the comparison

After the first week, run a live-vs-backtest trade comparison in RealTest. Export live trades from OrderClerk. Compare entry prices to what the backtest would have generated on the same dates. If fills are consistently worse than expected, fix the execution model before scaling up.

Mistake #12. You Don’t Have a Stop-Trading Number. You’ll Define It During a Drawdown.

If you don’t define your exit threshold before going live, you’ll define it emotionally during a drawdown, either too early (panic) or too late (denial).

Write down three numbers before day one:

1. Level where you reduce size by 50%

2. Level where you stop and review without emotion

3. Level where you stop completely and conduct a full strategy audit

These decisions, made with a clear head in advance, are the only reliable protection against making the worst possible call at the worst possible moment.

Mistake #13. You Skipped the Execution Dry Run

RealTest, OrderClerk, and TWS work perfectly in testing, until your first live trading day, when something unexpected always happens. Wrong order format. Account identifier mismatch. Time zone off by one. Now you’re troubleshooting under real-money pressure.

Run the full pipeline with paper trading before a single live dollar is committed. Know exactly what every confirmation screen looks like.

Mistake #14. You Launched Into a 40-VIX Market and Called It a Fair Test

Starting a mean-reversion strategy during a volatility spike is a stress test, not a launch. Your early results will be extreme and unrepresentative. The emotional relationship you form with the system in those first weeks, either “this is incredible” or “this is broken”, will distort your judgment for months.

If you launch into a crisis, acknowledge explicitly that the first 60 days of live data are not representative of long-run behavior before evaluating anything.

3. Active Trading Mistakes

This is where the real damage happens. The strategy is running. Every day brings new decisions, or the temptation to make decisions.

Mistake #15. You Skipped a Signal. You Called It Research. It Was Fear.

Your system generates a buy signal. The news looks bad.

“Just this once.”

Every time you override a signal, you’re making a discretionary decision your backtested edge doesn’t include. The trades you override most often are statistically the exact trades your mean-reversion edge is built to capture. Panic selloffs where sellers are exhausted and price snaps back.

Track every override. Log the date, symbol, signal type, and reason. After 90 days, run the theoretical P&L of all overridden trades. Traders who do this exercise stop overriding, not because someone told them to, but because the numbers make the argument better than any advice.

Mistake #16. You Stopped After Three Losses. All Three Were Normal.

Three consecutive losses feel different than the data suggested. The brain registers a pattern and the trader stops the strategy. But three consecutive losses are completely normal for any system with a 65 to 70% win rate.

The math, before you go live:

• 70% win rate > probability of 3 consecutive losses: 2.7%

• In 100 trades, you’ll hit that multiple times

• 5 consecutive losses: 0.24%, still not unusual over a year

Know these numbers before the first loss so they feel expected, not alarming.

Mistake #17. You Only Take the ‘Clean’ Signals. Your Edge No Longer Exists.

You take signals that look “right” on the chart and skip ones that look “risky.” The problem: your system doesn’t distinguish between clean and ugly. The edge is statistical across all qualifying signals. The ugly ones are part of the sample that produced the win rate.

Selectively taking a subset doesn’t improve the edge. It destroys it by removing the sample size the statistics depend on.

This is why automation matters. When a signal generates an order automatically, you don’t see the chart. You don’t decide. The edge remains intact because you’re not curating it.

Mistake #18. You’re Increasing Position Size After Five Wins. This Is the Worst Possible Time.

Five wins in a row. Confidence is high and you increase size for trade six.

Position size should be determined by your risk model (account equity, expected volatility, max position limits), not by your recent emotional state. Hot streaks are typically followed by reversion to the mean win rate. Now you’re oversized during the reversion period.

The correct response to a winning streak: verify the strategy is behaving within backtest parameters. Change nothing else.

Mistake #19. You’re Decreasing Position Size After Losses. This Is When You Need the Edge Most.

The mirror image of Mistake #18, and often more damaging. After five losses, you reduce size to “protect capital.” Statistically, a cold streak tends to revert toward normal performance. You’ve just reduced your exposure at the exact point where the edge is historically likely to reassert itself.

Unless your drawdown has reached a pre-defined review threshold, position sizing should not change based on recent win/loss sequences.

Mistake #20. You’re Checking P&L Every Hour. You’re Creating Noise, Not Information.

A position is down 3% intraday and you feel the need to do something. Your system was built on daily bars. Intraday fluctuations contain no information relevant to a daily-bar strategy.

Set one daily review time (after market close) and review closed trades only. Close the brokerage tab during market hours.

Mistake #21. You Paused During the Volatility Spike. You Paused During Your Best Month in 20 Years.

“Markets are crazy right now. I’ll pause the strategy until things calm down.”

This is the most expensive single mistake in systematic trading. High-volatility periods are often when mean-reversion edges are strongest, because dislocations are larger and reversions are faster. The evidence:

Every one of those periods was a moment when a discretionary trader might have paused. The traders who paused missed the best operating windows in a 25-year history.

Mistake #22. You Added a New Filter Mid-Run. You’ve Now Backtested Zero Days of That Change.

“I’m going to add a news sentiment check before taking signals.”

Adding filters to a live strategy is strategy redesign, not strategy management. You haven’t backtested the new filter. You’re changing the system based on weeks of live data, not years of historical data.

Any change should be developed separately, backtested fully on out-of-sample data, and launched as a new strategy, not grafted onto a running one.

Mistake #23. You Compared Today’s Return to SPY. That Number Is Meaningless.

Your strategy: +0.3% today. SPY: +1.8%. You feel like you’re losing.

This comparison means nothing on any individual day. A strategy with Sharpe 1.41 and -15.5% max drawdown is not built to win every day against SPY. It’s built to compound better over years with dramatically less downside. The correct comparison is annualized risk-adjusted return, not today’s P&L vs a benchmark that has no relationship to your strategy’s structure.

Mistake #24. One Bad Month. You Changed Parameters. Both Things Were Mistakes.

One bad month, and the first instinct is to tweak. But one month of live data is not statistically significant for a system built on 25 years of backtested data.

• One month = approximately 20 trading days

• 20 data points vs a sample of roughly 5,000

The math is not in favor of changing anything.

Mistake #25. You’re Still Manually Reviewing Signals Every Day. Every Day Is a Chance to Ruin This.

If you’re manually reviewing signals and placing orders every day, you’ve created a daily decision point. Every decision point is an opportunity for behavioral mistakes #15 through #24 to occur.

Full automation (RealTest generating signals, OrderClerk placing orders) removes most of these opportunities entirely. The five minutes of daily work becomes order review, not decision-making.

Paid subscribers also get access to the full vault:

• 53+ articles including

• 20+ with RealTest source code

Organized by category:

• Mean Reversion

• Trend Following

• Breakout

• Portfolio Construction

• Risk Management.

One subscription, everything published. → All articles and strategies here

Next:

Mistakes #26 to 32: Monitoring and Review

Mistakes #33 to 41: Drawdown and Crisis

Read more

Do you want to lower your drawdown (without giving up your returns) ?

It’s simple… to reduce your drawdown.

• Just lower your position size.

But at the same time, your profit decreases too.

• You do not want lower returns.

• You want the highest return possible.

Every systematic trader eventually hits the same wall.

Today, we’ll look at ways that help reduce your drawdowns and maximize upside.

The Full System: 4 Methods + Complete RealTest Code

Below I’m sharing all four methods in full, with the exact backtested code.

What you’ll get:

1. Method 1

2. Method 2 (+ RealTest Script)

3. Method 3 (+ RealTest Script)

4. Method 4 (+ RealTest Script)
   

100% actionable even for traders who are not using RealTest.

Read more

Should I even be using it at all?

Most traders think MACD is a trend-following indicator.

Here’s what you need to understand.

A 2026 academic paper tested MACD against a simple moving average on 70 global markets across 33 years of data.

The paper found that MACD does not outperform a single moving average for trend following.

I took the paper’s exact MACD formula and ran it on S&P 500 stocks. The result was a 73.8% win rate with an average winning trade lasting just 14 days.

MACD on trend following is fine if you enjoy the indicator, but don’t count on it to trend-follow on US equities.

So how do I even use MACD, or should I be using it at all?

Here’s what it’s actually doing instead.

What Is MACD and Why Does Everyone Use It?

MACD stands for Moving Average Convergence Divergence. It is one of the most popular trading indicators in the world.

The idea is simple. You take a fast moving average of price (usually the last 12 days) and a slow moving average (usually the last 26 days). When the fast one rises above the slow one, price is gaining momentum. That is your buy signal.

Most traders use MACD to identify trends. When a stock is trending upward, MACD signals that momentum is building. When it starts breaking down, MACD signals that the trend is weakening.

It sounds logical. The problem is that the research does not always support it.

The Paper That Made Us Question Everything

Sébastien Valeyre published a paper called “Breaking the Trend: How to Avoid Cherry-Picked Signals.”

His main argument was this. When you build a complex trading signal by combining multiple indicators, you risk accidentally tuning it to fit the past. The strategy looks amazing on historical data. But it only looks good because you unknowingly shaped it around data that no longer exists. This is called cherry-picking.

His conclusion was that a single, simple moving average is the best trend-following signal you can build. Adding more indicators, including MACD, does not help and can actually hurt.

He tested this on 70 markets globally including commodities, bonds, currencies, and stock indices from 1990 to 2023. The simple moving average got a Sharpe ratio of 1.245. MACD got 1.18.

A quick note on Sharpe ratio: it measures how much return you earn relative to the risk you take. Higher is better. A Sharpe of 1.0 is considered solid. Above 1.2 is excellent.

Simple won and MACD lost.

So I decided to try the paper’s exact MACD formula ourselves, but on S&P 500 stocks instead of global futures markets.

How I Ran the Test

I coded the paper’s MACD formula in a backtesting tool called RealTest.

A backtest means running a strategy on historical price data to see how it would have performed in the past. It is not a guarantee of future results, but it is the best tool I have for evaluating whether a strategy has a real edge.

I tested it on every S&P 500 stock from January 2000 to April 2026. Importantly, I included stocks that no longer exist, companies that went bankrupt or got acquired. Most backtesting tools skip these, which makes results look better than they really are. I did not do that.

The strategy holds up to 10 stocks at a time, sizes each position equally, and accounts for real trading fees and the small cost of not always getting the exact price you expected.

I also ran the paper’s recommended simple moving average at five different speeds, from fast (reacting quickly to price changes) to slow (smoothing out noise), to compare directly.

Right now,

Financial markets are not in good shape due to the war between Iran and the U.S.

This negative trend could continue for a long time because of higher gas prices and the expectation that rates will stay higher for longer, even if the war ends soon.

Thankfully, this is the best environment for mean-reversion strategies.

Here are some of our strategies that are currently benefiting from these market conditions.

• Mean Reversion Strategy (Last Weeks Return +3.21%)

• Weekly Buy The Dip Strategy (YTD Return +20%)

• Short Term MR (Larry Connor Style) Strategy (YTD +10.54%)

Back to the article…

What the Numbers Said

The simple moving average results were exactly what you would expect from a trend-following strategy.

The fastest version, reacting to 50 days of price history, earned a 4.3% annual return and won on only 39.8% of its trades. That is normal. Trend following loses more often than it wins. But the wins are big enough to make it worthwhile.

As I slowed down the moving average, results improved. The slowest version, using 180 days of price history, earned 9.62% per year with a Sharpe ratio of 0.61. Still a low win rate, around 38%. Still normal for trend following.

Now the MACD.

Annual return of 12.21%. Sharpe ratio of 0.74. Win rate of 73.8%.

The annual return and Sharpe beat the best simple moving average result. But the win rate makes no sense at all.

73.8% of trades were profitable. The average winning trade lasted 14 days.

That is not trend following. That is something else entirely.

Why 73.8% Win Rate Is the Wrong Number for a Trend Strategy

Here is the key difference between trend following and mean reversion, explained simply.

Trend following works by being wrong a lot. You buy a stock hoping it will keep rising. Most of the time it does not. You take a small loss and move on. But occasionally you catch a stock that climbs for months and that one trade covers all the losses. Professional trend-following funds typically win on 35 to 45% of their trades. The wins are just much bigger.

Mean reversion works the opposite way. Stocks that drop sharply tend to bounce back. You buy the dip and sell the recovery. You are right more often. But your gains are small and quick.

A 73.8% win rate with 14-day holds is clearly a mean reversion profile. Not trend following.

There is more. When the MACD strategy wins, it holds for 14 days on average. When it loses, it holds for 41 days. In real trend following, losers get cut quickly and winners run for months. Here it is completely backwards.

The strategy also made 3,139 trades over 26 years. The best simple moving average made just 221. That is 14 times more activity, with a completely different style.

Whatever this formula is doing on S&P 500 stocks, it is not following trends.

Why This Matters If You Trade US Stocks

If you have ever added MACD to a trend-following strategy hoping it would improve results and instead saw no improvement or even worse performance, this is likely why. On individual US stocks, MACD does not behave as a trend tool.

And if you decide to run this strategy yourself, understand that it will behave like a mean reversion strategy. It will not diversify your trend-following systems. It will move in line with your other short-term strategies.

Knowing what your strategy is actually doing is more important than knowing its backtest numbers.

Why US Stocks Behave Differently

Here is a well-known fact about individual US stocks. When a stock drops sharply over two to four weeks relative to the rest of the market, it has a statistically higher chance of bouncing back over the next couple of weeks. This is called short-term mean reversion.

It is one of the most studied and exploited patterns in US equity markets. Dozens of profitable strategies are built around it. The basic idea is simple: buy the dip in a healthy stock and sell the recovery.

The paper’s MACD formula was never designed to exploit this. But the negative 20-day weight accidentally does exactly that.

The formula prefers stocks that have pulled back recently. That is the contrarian short-term component at work. At the same time, the 120-day component acts as a quality check, making sure the stock is still in a solid medium-term uptrend before the strategy enters.

Put those two things together and you get a strategy that buys short-term weakness in stocks that are still fundamentally healthy. It holds for about two weeks, captures the bounce, and exits.

That is mean reversion with a trend filter on top. Not trend following.

That is where the 73.8% win rate comes from. Mean reversion strategies win frequently by design. The trend filter adds an extra layer by keeping the strategy out of damaged or falling stocks, pushing the win rate even higher.

The Full RealTest Code + Strategy Rules

This is the complete code. Survivorship-bias-free data, real trading fees, 0.1% slippage included, 10 positions at a time, S&P 500 stocks only, tested from 2000 to 2026.

Read more

You adjust the entry. No improvement and then you change the exit. Still the same. You add a regime filter and the backtest gets smoother but live trading doesn’t change.

The problem is not what you’re adjusting. It’s that you’re working on layer three when layer one is still broken.

Systematic trading strategies fail at specific points, in a predictable order. Not randomly. The same failure modes kept appearing in the same sequence across every strategy that didn’t survive. Seven of them.

These rules came from building the SetupAlpha strategy library over several years. The strategies that failed, failed at exactly one of these points, in roughly this order. Work through them sequentially. Each one clears the path for the next.

Rule 1: If You Can’t Explain Why a Strategy Works in One Sentence, Don’t Trade It

This is not a simplicity rule, its more like a fragility detector.

If you can explain your strategy in one sentence, you understand what it’s exploiting. And if you know what it’s exploiting, you know when that edge is gone. You have a way to diagnose underperformance that doesn’t involve re-optimizing everything from scratch.

For example:

One of the ETF rotation strategies I sell at SetupAlpha has a one-sentence explanation: every month, I buy the ETFs with the highest momentum over the past months.

You know why it works: momentum tends to persist over short windows. You know what regime it needs, trending markets where recent winners keep winning. You know when to be suspicious when the strategy underperforms in a clear trending market, something else is wrong not the logic itself.

You can see the full strategy at ETF Rotation Monthly Rebalance.

Now consider a strategy with this entry:

Buy when the 14-period RSI crosses above 52, the 50-day MA is above the 200-day MA, the stock is within 3% of a 20-day high, and volume is above its 20-day average etcetc.

Try writing one sentence explaining what market behavior that exploits. You can’t. Which means if it stops working, you have no idea what broke.

The test: write one sentence starting with “This strategy works because the market tends to...” If you can’t finish that sentence with something specific, go back and simplify until you can.

Rule 1 is the filter. The next six rules decide what survives it.

Rule 2: The Simple Version Almost Always Wins

When a strategy starts underperforming, the instinct is to add rules. A regime filter here, an extra condition there, a volatility screen somewhere else. After three rounds of this, the strategy has eight parameters and a suspiciously smooth equity curve.

The simpler version (the one you started with before the additions) usually performs better out-of-sample. Out-of-sample means data the strategy never saw during development the real test, not the rehearsal. Not always better. But often enough that the burden of proof should be on the complexity, not on the simplicity.

Before adding any rule or parameter, run both versions side by side for at least six months of out-of-sample data. If the complex version doesn’t clearly outperform, cut the addition. The simpler strategy will be easier to trust when it has a bad month, easier to diagnose when something changes, and less likely to be fitting noise instead of real market structure.

Rule 3: Mean Reversion and Trend Following Need Different Regime Filters

This is the mistake I see most often in portfolios of multiple strategies: the same SPY > SMA(200) filter applied to every strategy, regardless of type.

It works for trend following. For mean reversion, it often makes no sense at all. In many cases, no filter is the better answer.

Here’s why.

Mean reversion strategies profit from overreactions: stocks that drop sharply and snap back. The largest overreactions happen during selloffs. Selloffs happen below the 200-day moving average. If you filter out everything below SMA(200), you are filtering out the exact conditions where mean reversion has its strongest edge.

Trend following is the opposite. It needs sustained directional movement. The 200-day filter keeps trend strategies out of choppy, directionless markets, exactly the environment where trend strategies lose slowly on false signals.

Rule 4: Paper Trade With Live Data for Three Months Before Real Money

Not backtesting, not a walk-forward.

Live paper trading: real signals, real fills (simulated), real market hours, for at least three months.

Early in my trading career, I accidentally submitted a batch of orders incorrectly: wrong sizes on some positions. I was on a demo account, so nothing happened. The cost was zero. I fixed it, understood how it happened, and built a check into my process.

If that had been a live account, the loss would have been significant and immediate. Not because the strategy was wrong. Because order entry is an operational skill that breaks in ways backtests never show double submissions, incorrect position sizing from manual entry, data pipeline errors that generate phantom signals. These are not edge cases. They happen, especially in the first months of running a system.

The goal is not to confirm the strategy works. The backtest did that. The goal is to find the operational failures before they become expensive ones.

Next 3 rules are the most important in my opinion because:

• Rule 5: Don’t waste time on dead signals

• Rule 6: Build strategies that survive reality

• Rule 7: Don’t sabotage your own edge

Read more

Your stock drops 8%. You sell.

Next month it’s up 12%. Sound familiar?

There’s an academic paper that derives the mathematically optimal trend-following portfolio.

The formula it lands on says something interesting:

The exit doesn’t depend on the individual stock at all. Every position is controlled by one signal. The trend of the overall market.

I took that formula and turned it into a tradeable strategy.

In this article is will show you:

1. Entry, exit, ranking, all straight from the paper (+ paper link)

2. Full trading strategy rules

3. RealTest script

Here’s the strategy and what it returned.

The Core Idea

The market moves together. When stocks trend up, most of them trend up. When the market rolls over, most stocks follow. Individual stock signals are mostly noise on top of this shared trend.

So instead of watching 500 different exit signals on 500 different stocks, this strategy watches one thing: is the broad market still trending up?

How do you know? It’s simpler than you think, i’ll soon talk about it.

The entry side is just as clean. When the market trend turns positive, buy the least volatile stocks in the S&P 500. Why least volatile? Because in a trend-following context, low-volatility stocks give you the best risk-adjusted exposure. This comes from a concept called risk parity. You want each position contributing roughly equal risk to your portfolio. The simplest way to do that is to favor calm, steady movers over wild ones.

Why This Works

Think about it this way.

You buy a basket of S&P 500 stocks because the market is trending up. One of those stocks drops 8% on an earnings miss. What do you normally do? You sell it.because feels like the right call.

But if the overall market is still healthy and trending higher, that stock will probably recover. The shared market trend pulls everything along. By selling, you locked in a loss that would have turned into a gain.

The academic paper behind this strategy formalizes exactly this intuition. It shows that for a diversified basket of trend-following positions, the shared market trend is the dominant signal. Individual stock movements are mostly noise. Reacting to that noise with stop losses, individual exits, trailing stops, actually makes your portfolio worse. Not better.

The paper tested this on 47 futures across stocks, bonds, and currencies from 1985 to 2020 and achieved a Sharpe ratio of over 1.0. That’s strong for any trend-following system.

Results

Backtested on S&P 500 stocks, 2000–2026, $10,000 account:

A Sharpe of 0.75 won’t make headlines on its own. The paper achieved over 1.0. But they could go both long and short across stocks, bonds, and currencies. We’re limited to long-only equities. When the market drops, the paper’s strategy profits from shorts. Ours just goes to cash. That gap is structural, not a flaw in the approach.

Slippage and fees are significantly affecting the strategy.

What Makes This Different

Most trend-following strategies I see work like this calculate a trend for each stock, enter the best ones, exit each one when its own trend fades.

This strategy flips that completely. There is no per-stock signal. All 10 positions enter together and exit together, controlled by one market-level switch. The only thing that differs between stocks is their volatility ranking at entry.

Small losse, big winners.

The exit doesn’t depend on the individual stock at all. Every position is controlled by one signal. The trend of the overall market.

I took that formula and turned it into a tradeable strategy.

Now let’s talk about the paper & strategy itself, next:

1. The Paper

2. Strategy Rules

3. RealTest Script

Read more

Hey Trader!

Are you aware that markets are competitive but not perfectly efficient?

Certain patterns keep appearing because different market participants behave in predictable ways.

• Large funds follow rules

• Investors react emotionally to events

• And many traders repeat the same mistakes.

These behaviors create repeatable opportunities, often called “trading edges”.

For systematic traders, the goal is simple: identify patterns that have worked historically, test them properly, and apply them with clear rules.

In this article we catalogue 60+ market edges used by traders, organized into several categories. For each one we outline:

• Why it exists

• How traders access the data or signal

• How it is typically used in practice

1. Let’s start with the broker / execution edges

Before strategy even matters, your broker and execution setup can be an edge or a hidden tax that kills your returns. Here’s every broker and execution edge you can exploit:

Pro tip: Retail traders never optimize their broker setup. Switching from a high-commission broker to IBKR Pro with smart order routing alone can save you 0.5–1% per year on an active account. Over a decade, that's tens of thousands of dollars.

2. Factor Edges (Academically Proven)

Decades of academic research have identified persistent return premiums "factors" that explain why certain stocks outperform.

These are not some X trader opinions, they're backed by thousands of peer-reviewed papers.

Where I find the best academic edges?

I personally spend a lot of time reading papers on SSRN and arXiv.

These are goldmines for discovering new factor premiums, anomalies, and trading strategies that most retail traders have never heard of. The key is to look for papers with large sample sizes, out-of-sample testing, and results that survive transaction costs. If a paper only works with zero slippage and zero commissions it's not a real edge(not for me). Focus on the ones that are robust enough to trade in the real world.

3. Information Edges

An information edge means you know something (legally!) before the market prices it in. This is arguably the most powerful edge category but also the hardest to build.

Real-world example: One thing I've learned is that the best information edges come from combining domain expertise with publicly available data. Here's a great example:

A friend of mine works in the aviation industry and exclusively trades airline and aviation stocks. His secret sauce? FlightRadar24. He tracks specific airlines' aircraft how many are flying, how many are sitting parked on the ground, route changes, fleet utilization rates. Because he understands the industry deeply, he can interpret this data in ways a generalist never could.

When he sees an airline's fleet utilization dropping weeks before their earnings call while the Street still expects strong numbers he has an information edge. It's completely legal, completely public, and extremely powerful because he has the domain knowledge to make sense of it.

4. Strategy / Process Edges

This is where most traders think they have an edge but few actually do.

A real strategy edge is not 5 magic indicators, its usually a repeatable, tested process that exploits a genuine market behavior.

This is where SetupAlpha comes in.
At setupalpha.com and in this Substack, we build and share real, tested trading strategies, not hypothetical ideas or cherry-picked backtests. Every strategy we publish is:

• Out-of-sample tested

• Walk-forward validated

• Stress tested across different market regimes

• Built with full RealTest code you can use (+ Rules step-by-step)

Here are two examples of Substack strategies we've published with full rules and code:

Subscribe now

5. Structural Edges

Structural edges come from the way markets are organized rules, mechanics, and participant behavior that create repeatable inefficiencies.

These are often the most persistent edges because they're built into the market's “DNA”.

Here's a structural edge many people don't talk about: crypto markets have massive wicks those huge candle tails where price crashes 10–20% and snaps back within minutes.

Why does this happen? Because crypto allows extremely high leverage. Most crypto traders max out their margin 10x, 25x, even 100x. When the market drops, it triggers a cascade of forced liquidations. Brokers and exchanges are forced to close all these leveraged positions at market price, which pushes the price even lower. But then once the liquidations are done there are no more sellers, and price bounces right back.

If you understand this mechanic, you can place limit orders at deep discount levels during high-leverage periods and let the liquidation cascade come to you. Combine this with monitoring liquidation data from tools like Coinglass to see where the leverage is stacking up and you have a structural edge that most traders don't even know exists.

6. Volatility Edges

Volatility is one of the most mispriced things in financial markets.

If you understand how to trade it, the edge can be enormous.

Warning: Volatility edges can blow up spectacularly if you don't manage risk.

The VRP (selling options) can work 85% of the time but the 15% can wipe you out if you're over-leveraged. Always size these strategies conservatively and have a hard stop for maximum loss per trade and per month.

7. Time-Based Edges (Calendar Effects)

Markets have persistent time-based patterns.

Some are well-known, some are still exploitable.

Here are the calendar effects that academic and quantitative research has documented:

Calendar edges are weakening but they haven't disappeared.

The key is combining them with other edges (strategy + timing).

A mean-reversion strategy that also incorporates turn-of-month timing can significantly outperform the same strategy traded any day of the month. Stack your edges.

Putting It All Together

Most traders have zero edges. Some have one. The best traders stack multiple edges from different categories.

A trader who combines a strategy edge (tested, robust system) with a broker edge (low commissions, smart routing) and a time-based edge (turn-of-month timing) and an information edge (domain expertise in one sector) is operating on a completely different level than the guy following stock picks on X/Twitter.

Your homework after reading this:

1. Audit your current edges. Write down every edge you have. If you can’t name at least two you probably don’t have one.

2. Pick one new category from this article and start building an edge there this month.

3. Stack, don’t switch. Don’t abandon what works. Add new edges on top of your existing ones. I always try to integrate diverse edges into my strategies.

Now you have the map. Go build yours.

Next week, we're going deep.

We'll be publishing a full deep-dive into one major trading strategy complete rules, backtest results, RealTest code, and everything you need to trade it yourself.

Don't miss it. Subscribe now so it lands in your inbox the moment it drops.

Subscribe now

When SPY drops, you probably rotate into defensive sectors healthcare utilities. consumer staples. That’s the move and it usually feels responsible (if you are not trading algorithmically yet) but it’s not protecting you.

I ran

• 88 backtests across

• 22 assets and

• 21 years of data.

The sectors you rotate into during a crash give you the same annual return as the best refuge assets in the test. But with two to three times the drawdown.

Holding Financials (XLF) when SPY is in crash mode: 7.5% annual return, -76% maximum drawdown.

The best-ranked asset in this test produces a similar annual return with less than a third of that drawdown.

Same return and triple the drawdown. You are not rotating to safety because you are paying three times the drawdown for nothing.

• Healthcare during crashes: -44% average drawdown.

• Utilities: -48%. The sectors that feel defensive are some of the worst assets you can hold when it matters most.

The problem is you probably don’t know exactly which ones destroy capital and which ones actually hold up.

Retail traders usually find out during the crash, not before it. That’s too late.

This article:

• The full risk-adjusted ranking of all 22 assets, worst to best

• Which specific sectors to avoid and rotate out of when SPY drops

• The complete RealTest code to run this yourself

Read more

In 1998, Long-Term Capital Management had two Nobel Prize winners on its board, four billion dollars in capital, and what many considered the most sophisticated trading strategy ever built.

In less than four months, they lost it all.

Not because their strategy was bad. it had a genuine statistical edge. The problem was portfolio. It was their only strategy. When the Russian debt crisis (1998) hit and their single strategy faced conditions it had never encountered, there was nothing else to absorb the blow.

LTCM (Long-Term Capital Management) is not an outlier.

Amaranth (The Amaranth Case Study) lost six billion in a single week on one natural gas strategy.

Tiger Management shut down when their concentrated value approach collided with the tech bubble.

Meanwhile, multi-strategy platforms like Citadel and Millennium not only survived these same periods, they continued compounding. Millennium has produced positive returns in twenty-eight of its thirty years.

The Only Free Lunch in Finance

The principle behind that system is one of the most well-documented concepts in all of finance. Markowitz won a Nobel Prize for it in 1952. Ray Dalio calls it the Holy Grail of investing.

The majority of retail traders still allocate everything to a single strategy.

Consider what happens when you combine two structurally different strategies. Say your first is mean reversion buying stocks that dropped sharply and tend to snap back. Your second is a breakout system buying stocks making new highs in strong trends.

These strategies profit in different market conditions. When mean reversion struggles in a sustained trend, breakout does well. When breakout grinds out losses in a choppy, sideways market, mean reversion thrives.

Combined: the highs are slightly lower, but the lows are much higher. Drawdowns become shallower and shorter because while one strategy dips, the other cushions the portfolio.

Markowitz called this the only free lunch in finance, reducing risk without sacrificing expected return.

Dalio quantified it precisely:

• One strategy: baseline risk

• Add one uncorrelated strategy: risk drops ~30%, same expected return

• Add a third, fourth, fifth: risk drops 60%

• Fifteen to twenty uncorrelated strategies: risk drops 80%

How the Institutions Structure This

Citadel, Millennium, and Point72 run what is called the “pod model.” Each pod is an independent team running its own strategy equity long-short, macro, statistical arbitrage, credit, commodities… all under one roof with centralized risk management.

Millennium enforces strict drawdown limits. A pod hits 5% drawdown: capital is halved. Hits 7.5%: pod is shut down entirely.

The result is higher risk-adjusted returns, lower volatility, and smaller drawdowns at the firm level even when individual pods are struggling.

The question is whether a retail trader can apply the same principles. The math does not care about account size.

What the Data Shows

The following is a concrete demonstration using real strategy backtests. All numbers are from historical data.

Starting point: $100,000, monthly rebalancing.

Portfolio 1: Single Strategy

100% allocation to a single ETF Rotation system.

• CAGR: 10.95%

• Sharpe ratio: 0.90

• Maximum drawdown: −18.6%

When your only strategy is underwater, 100% of your capital is underwater.

During the COVID crash, this single strategy dropped 8.8% every shock hits your full capital.

Good strategy but fragile portfolio.

Portfolio 2: Two Strategies

Adding a weekly pullback mean reversion strategy alongside the ETF rotation system. Fifty-fifty split.

• CAGR: 13.9% (up from 10%)

• Sharpe ratio: 1.14 (up from 0.9)

• Maximum drawdown: −15.29% (compressed from −18.60%)

The equity curve is noticeably smoother. The correlation between these two strategies is 0.42 structurally different enough that their rough patches rarely overlap.

This is the free lunch beginning to appear.

Portfolio 3: Four Strategies

Adding a short-selling strategy and a second mean reversion system. Equal weights across all four.

• CAGR: broadly maintained

• Sharpe ratio: 1.70 (up from 0.9 with one strategy)

• Maximum drawdown: −11.27% (from −18.5%)

• Calmar ratio: 1.40 (from 0.59)

The Rebalancing Question

One additional variable worth examining: does rebalancing matter?

Without rebalancing, the maximum drawdown actually drops to −8.88% (from −11.27%), and Sharpe rises to 1.78 (from 1.70).

Those numbers look better buut that is not the full picture.

Without rebalancing, whatever performed best takes over an increasing share of the portfolio and you gradually lose the diversification you built. The portfolio drifts back toward concentration without you noticing.

Monthly/Quarterly or annual rebalancing maintains structure. It systematically trims positions that have grown overweight and adds to those that have lagged which is mechanically buying low and selling high at the portfolio level, automatically, without discretion.

Three Rules for Building This Yourself

1. Think in portfolios, not individual strategies.

Blue = all strategies equally weighted.
Red = Single SPY mean reversion one strategy.
SPY = benchmark.

The evaluation question is never “is this strategy good?” It is “what does this strategy do to my portfolio?” A strategy that improves Sharpe and reduces drawdown at the portfolio level earns its allocation regardless of standalone metrics.

2. Aim for four to ten strategies.

Below four, a single bad strategy drags the whole portfolio. Above fifteen, the marginal benefit of adding another strategy shrinks rapidly. The biggest risk reduction happens going from one strategy to five or six.

3. Rebalance on a schedule.

Monthly or quarterly or annual pick one and commit. Without it, your portfolio drifts back toward the same concentration you started with.

On Uploading Your Own Data

If you have your own strategy, any system and you have its equity curve as a CSV with two columns (Date, Equity), you can load it into the same analysis.

Run the correlation matrix against other strategies. See what it does to combined drawdown. Measure whether it actually adds to portfolio-level Sharpe or just adds noise.

Subscribe now

The Actual Holy Grail

There is a recurring search in systematic trading for the single perfect strategy the one system that works in all conditions, never has drawdowns, and compounds forever.

That system does not exist. Every strategy has a market environment where it struggles. The institutions that have compounded for decades have not found a better individual strategy they have built a better architecture around multiple strategies.

The Holy Grail is not one perfect system. It is the systematic coordination of multiple independent systems into a unified, resilient portfolio.

The math behind this has been published, replicated, and validated since 1952. The only thing missing is applying it.

The portfolio construction tool referenced in this article is free to use.

Last week’s strategy ran a Sharpe of 1.07, a max drawdown of -13.64%, and a win rate just under 70%. Those numbers held up well over 11 years of data.

But I kept running the test. One question wouldn’t go away: when the system steps aside from equities, the capital sits in cash. Cash earns nothing. Is there a better destination?

I added one extra asset. Same signal, no new parameters.

Max drawdown dropped from -13.64% to -12.16%. MAR ratio annual return divided by max drawdown improved 25%, from 0.96 to 1.20. Annual return went from 13.13% to 14.57%. Net profit on a $20,000 account increased by $11,658 over the same test period.

This article shows the strategy, why it works, and gives you the full updated RealTest code + strategy rules.

If you haven’t read last week’s article yet, start there this builds directly on it.

Ray Dalio stood at the World Economic Forum in Davos in January 2020 and said, on live television:

“Cash is trash. Get out of cash. There’s still a lot of money in cash.”

Two months later, SPY was down 34%. And something else was up double digits.

People filed Dalio’s quote as a January hot take and moved on. But he wasn’t talking about market timing. He was pointing at a capital problem every systematic trader eventually runs into.

When a system steps aside from equities, the capital has to go somewhere. Most of the time it defaults to cash. And cash earns nothing while the market keeps rotating.

When the market is in a risk-off environment where is the money actually going?

Not “what’s the safe place to hide.” Where is the active rotation happening?

A Paper

In 2006, Meb Faber published a white paper called “A Quantitative Approach to Tactical Asset Allocation.” It became one of the most downloaded investment research papers ever written.

The central finding was this: a simple, rule-based system that rotates between asset classes based on trend signals can produce equity-like returns with bond-like drawdowns. The key word is rotates.

Faber’s work showed that different asset classes take turns leading during different macro environments. When one is in a downtrend, another is almost always in an uptrend. Capital has somewhere to go, you just have to build the system to follow it.

It shows up in the actual historical data, cycle after cycle.

In the 2020 COVID selloff, while U.S. equities were collapsing, IEF the 7-10 year Treasury ETF posted a full-year return of over 10%. TLT, the longer-duration version, returned 17.9% for the year. When equities panic, capital does not disappear. It rotates into safety.

The 2022 rate cycle broke this pattern. Bonds got hit alongside equities. That’s a real caveat and one worth taking seriously. But across the full window of data we’re working with 2015 to today the rotation holds.

The System Already Knows

Here’s what struck me when I ran this next backtest.

The system we built last week already has the signal. It identifies when the macro environment is risk-off. It knows, with reasonable historical accuracy, when conditions are deteriorating for equities.

That same signal also happens to describe exactly the environment where the flight-to-safety rotation has historically worked.

So the question became: what if we gave the system a destination, instead of a waiting room?

I backtested it. Added one asset.

The MAR ratio return divided by max drawdown, the number that tells you how hard the capital is working relative to the pain it takes went from 0.96 to 1.20. That is a 25% improvement in capital efficiency. In the backtest data, 2015 to February 2026.

Max drawdown dropped. Annual return went up. Net profit on a $20,000 account increased by over $11,600 across the test period.

One asset, the same signal but a fundamentally different use of the time the system previously spent doing nothing.

If you’re building RealTest systematic strategies and want more of this kind of approach, the strategies section on SetupAlpha.com is worth exploring.

In this article, you will unlock:

• The asset: What activates when the system steps aside from equities and why the logic behind it is grounded in decades of institutional capital flow data

• The complete RealTest code: Full upgrade to last week’s script, copy-paste ready

• Trading Strategy rules (Step-By-Step)

Read more

Imagine hiring a security guard for your home who only works from 9:30 AM to 4:00 PM. If a fire starts at midnight, or a burglar breaks in at 3:00 AM, he’s not there. He’s “off the clock.”

This is exactly how the US Equity market functions. It is open for a mere 6.5 hours (tradeable liquidity) a day, five days a week. That leaves 17.5 hours each day plus entire weekends where your capital is locked in a burning building with no exit.

We do not live in a 9-to-5 world. Geopolitical conflicts, currency devaluations, and banking failures do not wait for the Opening Bell.

Information travels instantly.

Thirty years ago, news of a banking crisis might have taken days to fully impact global sentiment. Today, a single tweet or a leaked report spreads to millions of people in seconds. The reaction function of the human herd has accelerated to light speed, yet the stock market remains stuck in a 19th-century operating window.

• When war breaks out in Europe, it’s often 2:00 AM in New York.

• When an Asian bank collapses, Wall Street is asleep.

If you are holding SPY (S&P 500) blindly, you are exposed to “Gap Risk.” You go to bed with your portfolio intact and wake up to a market that has already repriced 5% lower before you can even click “Sell.”

The “Safe” Assets Are Failing You

For decades, the standard advice was simple: “Buy Bonds. Diversify.”

But as the WisdomTree “Bitcoin Correlations” research paper demonstrates, the rules have changed. In 2022, we saw a correlation breakdown that terrified institutional investors: Stocks and Bonds fell together. The correlation coefficient between the S&P 500 and U.S. Treasuries approached positive 1.0.

The “flight to safety” trade is broken because in a liquidity crisis, everything correlates to 1. Investors sell whatever is liquid to cover margin calls. Relying on traditional diversification is no longer a hedge, its a liability.

Testing the “Canary” Hypothesis

The theory is compelling because Bitcoin trades 24/7 and reacts instantly to global liquidity, it should signal danger before the S&P 500 opens.

Research from Delphi Digital supports this, suggesting Bitcoin often tops weeks before equities (warning of the 2022 crash nearly two months in advance). But theory isn’t enough. We need hard data.

So, I decided to run an experiment.

I ran a series of backtests in RealTest to answer one question: Can a specific signal from another market efficiently rotate us out of the S&P 500 before a crash?

I tested the most common institutional settings to find the truth.

The results were fascinating. Most common settings failed miserably, lagging the market or getting chopped to pieces. But one specific configuration stabilized the signal, doubling the Sharpe Ratio from 0.50 to 1.07 and significantly reducing maximum drawdown.

Next, I’ll share the exact:

1. Strategy rules (Step-By-Step)

2. Experiment results

3. RealTest code

And the specific anomaly that turned this theoretical correlation into a backtested edge.

Read more

Run the year-deletion test on most of them. Half collapse.

Not because the research is bad. Because aggregate metrics hide what individual years reveal. A strategy can produce a 2.4 Sharpe across 15 years and still be carrying a single event that generated 60% of the total return. Remove that year, and the Sharpe drops below 1. The paper does not mention this. The authors likely never tested it.

This is not a problem unique to academia. It is structural to how backtests are evaluated.

There is a test that takes less than 60 seconds. It exposes the dependency that aggregate metrics cannot see.

Why Standard Risk Metrics Miss This

Sharpe, Sortino, CAR/MDD etc none of these metrics reveal how returns were structured.

They confirm that returns existed. They do not expose the architecture underneath: where alpha was concentrated, what regime generated it, and whether that regime has any reason to repeat.

A strategy with a 1.2 Sharpe and a hidden dependency on a single market dislocation is not a robust strategy. It is a one-event position that happened to backtest across 10 years without revealing itself.

The Sharpe does not see it. The drawdown does not see it. The win rate does not see it.

One test does.

The Test

Open the backtest. Pull up the yearly performance table. Identify the year with the highest contribution to overall equity. Delete it.

In RealTest:

EntrySkip: Year=2020

Run the backtest again. Observe the CAGR, drawdown, and equity curve as if that year never occurred.

This is the strategy without its best witness.

Reading the Output

The numbers will drop. The common response: “Of course, I removed a good year.” That is rationalization, not analysis.

The correct question is not whether the numbers dropped. It is by how much they should drop.

A system generating 17% CAGR over 10 years with one year removed should still produce something real. If the residual is 3%, this is not a 10-year strategy. It is a 1-year strategy that avoided ruin for 9 years. A fundamentally different instrument.

The Questions This Exposes

The best year is identified. But the diagnostic has barely started.

Does the entry logic align with a macro regime that has not existed since that year? Did the exits function because of a volatility environment that permanently shifted after 2022? Is the edge in the system itself or is the system waiting, year after year, for conditions that may not return?

There is a deeper version of this test. One that separates structural alpha from situational alpha. The difference between a system traded with conviction and one traded on hope.

That is what the second half builds.

Read more

“Trend following is dead.”

We hear this refrain every year. And reliably, simple momentum systems manage to capture the outliers that drive 80% of long-term portfolio returns.

But for the systematic trader, the challenge is rarely identifying the trend. The challenge is the Execution. Specifically: what does your system do when it receives a valid signal but your portfolio is full?

This week I want to talk about one of the most critical and misunderstood aspects of portfolio management: Path Dependency.

The Scenario: Missing the Train

Imagine a classic trend-following strategy say, a 20-day breakout or a moving average crossover.

1. Month 1: Asset A hits a buy signal.

2. The Constraint: Your portfolio is already 100% invested. Your system forces a “Pass.”

3. Month 2: Asset A is now up another 10%. You finally sell an old position and have cash available.

The Dilemma: Do you buy this “late” trend? Or has the train left the station?

The “Loss of Momentum” Fallacy

Intuitively, most discretionary traders shy away from the second entry.

• “I missed the good price.”

• “The risk-reward is ruined.”

This is Anchoring Bias anchoring your value perception to the price you could have had rather than the price the market offers today.

A Real-World Example: ETF Rotation

The best way to understand this is to look at strategies that force you to ignore entry price and focus entirely on Relative Strength State.

Take our ETF Rotation Monthly Rebalance Strategy.

This system doesn’t care if it “missed” the bottom of the Nasdaq in January. It simply asks one question on the first of every month: “What is the strongest asset class right now?”

• If Equities (QLD) are leading, it holds Equities.

• If Bonds (TLT) are performing, it shifts to Bonds.

• If Crypto (IBIT) are outliers, it rotates there.

By design, this strategy is almost always “late” to the initial turn. It waits for the trend to prove itself on a monthly timeframe. And yet, because it systematically rotates into what is working (and cuts what is dying), it avoids the catastrophic drawdowns of buy-and-hold while capturing the bulk of major bull runs.

The lesson here is simple: The State of the trend (is it leading?) matters infinitely more than the Event of the breakout (did I catch the first green candle?).

But is there a mathematical penalty for entering late?

Here is the data-driven breakdown of why "perfect" entries are a mathematical anchor on your portfolio and the three specific "Late Entry" signals that actually outperform the breakout.

The Quantitative Reality: Trade vs. Portfolio

Read more