This suggests an idea: can I use it as a signal? If, for example, I see the market at 0.55/0.45 a few seconds before the candle open (a 5-cent skew toward Up), I can enter a position in that direction on a crypto exchange, and hold it until the candle closes. I should win 55% of the time — and the previous article says I really would, but do I have an edge? Let’s find out.

These questions will help us understand if this idea is worth something. Try answering them yourself before you continue reading:

1. How much does an ordinary candle move (5m, 15m or 1h)?

2. Are wins and losses symmetrical?

3. Does it survive fees?

   Thanks for reading Viktoria's Crypto Trading Lab! Subscribe for free to receive new posts and support my work.

How much does an ordinary candle move?

The timeframes that I’m interested in are 5m, 15m and 1h, since those have the largest Polymarket datasets. For average absolute candle returns, I used Binance BTC/USDT spot 1-minute data from 2020 to 2026, resampled to each timeframe. The results, grouped by year, are shown below in a plot and a table:

^{*2026 to 1 Aug}

Look how well the square root of time rule holds for every year. A 1h candle is four times longer than a 15m one and moves about twice as far (in 2026, 0.32% vs 0.16%). Which gives us an intuition. If short candles are too small to pay fees, trade longer ones. We’ll see in the next section if this is true.

As you can see, 2021 was the most volatile year, and since then average candle size has fallen by half. It is important to use the latest data rather than the full dataset so you are not biased by earlier, more volatile periods. This data is already enough for a quick calculation of potential edge and breakeven fees using the simple formula:

Edge = (Win Rate × Average Win) - (Loss Rate × Average Loss)

Assuming symmetric wins and losses:

Edge = Average Move × (2 × Win Rate - 1)

With a 55% win rate (probability skew) and 2026 average move we get these potential edges:

Now compare those breakeven numbers to what exchanges offers. On perpetuals, Binance is 0.02% maker and 0.05% taker; Hyperliquid is 0.015% maker and 0.045% taker. Only the 1h timeframe with the largest candle size is right at the edge of the breakeven fee. So the preliminary verdict is that the strategy isn’t viable under the assumption of a symmetrical win loss ratio. But let’s go further and see whether this assumption ever holds.

Are wins and losses symmetrical?

Signal. For 5m and 15m markets I use the last Polymarket trade before the candle open, so to get the signal as close to open as possible (median age: 1s and 2s). For 1h markets, last minute trades are less frequent, so I use the mid price from one-minute orderbook snapshots instead (bid/ask average). A trade is taken only when the signal skew exceeds a threshold away from 50/50: long if the Up price is higher, short if Down is. Entry is the open of the candle’s first minute and exit is the Close of its last, using Binance historical 1min klines.

How to read the tables. Each row is cumulative and is a trading rule. ≥ 3¢ means: trade only when the market is at least 3 cents off 50/50 in either direction (for example 0.53/0.47 take a trade, 0.51/0.49 don’t). So the ≥ 1¢ row includes everything in ≥ 3¢ and ≥ 5¢, plus weaker markets. Edge is the average gross return per trade, before fees. Breakeven fee is that edge divided by two.

_5m

_{Signal: last trade before open, median 1 sec | Mar 2026 - July 2026 | 43,864 markets}

_15m

_{Signal: last trade before open, median 2 sec | Nov 2025 - July 2026 | 25,907 markets}

_1h

_{Signal: order-book mid, median 54 sec before open | Nov 2025 - July 2026 | 6,252 markets}

The results show something interesting.

The direction is right. The same finding as in the previous research. The realized win rate is very close to the prediction on all timeframes. And it rises as the skew grows.

But wins are 10-15% smaller than losses. When the prediction is right, it yields less return than when it is wrong across all timeframes and at every threshold. The charts below use 15m as the example, but 5m and 1h look similar.

That makes the edge much worse than in the symmetric case. Compare the actual edge per trade with what it would have been under the symmetric assumption from the previous section. For example, at 15m ≥3¢ the symmetric edge would be 0.0159%, while the real one is just 0.0054% — 66% less.

Another finding is that on higher timeframes the potential edge doesn’t increase much. On 1h it greatly improves at ≥5¢ threshold but there are just 451 markets, so it shouldn’t be taken as a great finding.

Does it survive fees?

We’ve almost answered that question above. The strategy can only survive with near-zero costs. That could be realistic with higher VIP tiers or zero-fee promotions. But, keep in mind that orders can still suffer from slippage and spread, that will reduce even tiny edge more.

Still, let’s look at the performance under hypothetical zero fees. Below are the equity curves for 5m and 15m with a skew threshold ≥3¢.

5m, zero-fee

15m, zero-fee

Introducing any fee makes even the best performance look untradable. With maker fee of 0.015% per side on 5m timeframe with ≥3¢ skew, turns +115% into −94%.

5m, 0.015% fee

Summary

Polymarket predicts the direction correctly, and the win rate is close to the predicted one. But the strategy is hard to trade. First, candle returns are quite small relative to fees. Second, when we’re right we make less than we lose when we’re wrong. Sign is not magnitude. The edge per trade is tradable only at zero cost.

Run it yourself

No need to install anything. Just run it on the Google Colab. The full per market dataset is published as a CSV file, every Polymarket signal with the matching Binance entry and exit, all three timeframes. Just change three numbers, re-run and get your own equity curve and drawdown:

TF        = '15m'   # '5m' | '15m' | '1h'
THRESHOLD = 0.03    # min skew to take a trade |price_up − 0.50| 
FEE_PCT   = 0.00    # fee per side

Method notes

Signals come from Polymarket's free public APIs: Gamma for market metadata, the Data API trade feed, and CLOB one-minute price history); entry and exit prices are Binance spot BTC/USDT 1-minute klines: 43,864 markets at 5m (Mar 1 – Jul 31 2026), 25,907 at 15m and 6,252 at 1h (both Nov 1 2025 – Jul 31 2026). For 5m and 15m the signal is the last trade before the open, at most 30 seconds stale and half-spread corrected to approximate the mid between bid and ask; for 1h it is the order-book mid at T−1m, at most 120 seconds stale, because slow markets often have no last-minute trade, even if the book is quoted. Entry is the open of the candle’s first minute and exit is the close of its last minute. All returns are gross of funding and slippage, and assume fills at the exact open and close.

Nothing here is trading advice.

Resources

• GitHub Repo — dataset, notebooks

• Run the notebook in Colab — one click, no setup

• Polymarket — prediction market platform discussed (referral link)

• Polymarket API docs — the free public APIs this data comes from

_{Polymarket links in this post are referral links}

Polymarket’s Up or Down crypto markets ask one question: will the candle close be higher than its open? In theory, before the candle opens, the probability of it going up or down should be close to a 50/50 coin flip. In practice, it is quite common to see prices fluctuate, with one side at $0.48, $0.46, or even at $0.44, as if the crowd already knew the likely direction. Betting and trading start long before the open price is known.

This skew made me curious to study three questions:

1. When does the skew appear? Early, or only just before the period opens?

2. What range is typical? Can it exceed 60/40?

3. Prediction vs Reality. If the market says 55% Up, does Up win 55% of the time?

Data

I was mostly interested in the 15-minute BTC market, with checks on other timeframes as well. From the Polymarket endpoint, I pulled order-book history from November 2025 through July 2026 — one-minute snapshots of mid prices (bid/ask average). I used only the Up mid prices, because the Down mid should always be 1 − Up.

After filtering out bad data (empty history, missing pre-open prices, price disagreements, stale snapshots, and so on), about 23,700 markets remained with a clean snapshot 1 minute before open. Of those 19,439 had clean snapshots at all seven offsets before open: 60, 30, 15, 10, 5, 2, and 1 minute. I used this balanced panel for the first question, so the market mix is identical across offsets, and the full 23,700 for the rest analysis.

When does the skew appear?

Here is the answer in one picture. Each vertical slice is the distribution of pre-open prices at that offset, with quantiles shown. An hour before open, the price reflects a coin flip, 90% of prices sit between 0.495 - 0.505. Skew only shows up closer to the period open, mainly in the last 5-10 minutes, and the range is widest at 1 minute out. This makes sense, whatever models traders use, the most relevant information arrives near open, not an hour before.

One small curiosity: the median mid sits at 0.505 rather than 0.5 at every offset. Part of this is mechanical: with 1¢ ticks and a 1¢ spread, the mid can only be 0.495 or 0.505 — but books lean toward the higher, Up-premium configuration more often. This plot also gives us a preview of the second question.

What range is typical?

Here I looked at prices 1 minute before open. The short answer to "can it exceed 60/40?" is: almost never. Even at that point, 90% of markets stay within ±5 cents of a coin flip (0.45 - 0.55), and 99.9% of prices are between 0.4 and 0.6. The histogram below peaks at 0.49 - 0.51 as expected, but more than 65% of markets sit outside that peak — and those are the main interest of this study.

Prediction vs Reality

That brings us to the third question: should you trust the prices? I bin the implied probabilities (Up mid-prices from the snapshot 1 minute before open) into 1-cent bins, keeping at least 1,000 markets per bin and putting leftovers into the edge bins. The x-axis is each bin’s mean Up price, while the realized Up frequency is on the y-axis. Dot size scales with bin count, the 45° line is shown as the reference for ideal prediction, and the whiskers are 95% confidence intervals:

The realized outcomes follow the reference line closely. From bins below 0.45 to above 0.55, realized Up frequency matches the quoted price to within about a cent in nearly every bin. Markets priced near 56.5¢ resolved Up 57% of the time; markets priced near 47.4¢ resolved Up 46.8% of the time. Only one bin (0.45-0.46, shown as red dot) has a confidence interval that falls statistically off the diagonal, which is about what you’d expect by chance across twelve bins.

Here is the complete table of the bin values:

What does this mean?

So, does Polymarket know which way BTC will move? A little, and that little is already in the price. These pre-open prices aren’t random noise. On average, they are a good forecast of Up vs Down in a short range of ±5 cents. If you were hoping to trade against the pre-open skew, thinking that true odds are 50/50 and the skew is an a opportunity, the data shows there is no statistical edge in that. And keep in mind that everything above uses order book mids, not executable fills. A taker crossing the spread also pays a fee, so a mispricing this small can easily turn into an unprofitable trade. Whether passive maker orders, which earn rebates instead, can capture any edge is a different question.

Still, I hope this gave you some intuition and new ideas for research. It gave me some. I’ll unpack them in later posts.

Stay tuned!

Method notes: 23,909 raw BTC 15 min Up/Down markets, Nov 1 2025 – Jul 8 2026 (249 days). All prices here are order-book mids, not executable fills. Order books snapshotted at fixed offsets before each open; markets with empty histories, stale quotes (>60s), or Up/Down books disagreeing by more than a cent excluded (~0.9% of the sample). The skew-vs-time analysis uses a balanced panel of 19,439 markets clean at all seven offsets, so the market mix is identical across offsets; calibration uses all 23,700 markets with a clean 1-minute snapshot, binned in 1-cent buckets of ≥1,000 markets, with Wilson 95% intervals.

Other timeframes

Here’s a sneak peak of the same analysis for 5-minute and 1-hour markets. The 1-hour series uses 2 cents bins because there is less data. The plots show a similar picture: the outcomes follow the reference line quite closely. More data and detailed results on Github.

Resources

• GitHub Repo — all datasets (full parquet + cleaned CSV), notebooks, and the scripts to re-download everything from Polymarket's public APIs.

• Github Results for every timeframe — all plots and tables for 5m, 15m and 1h

• Run the notebooks in Colab — one click, no setup

• Polymarket — Prediction market platform discussed

• Polymarket API docs — the free public APIs (Gamma metadata, CLOB price history) this data comes from

• I'll explain how funding rates work

• Share the custom Python script I developed to find trading opportunities

• Reveal the results of my trades

This is all for fun and learning, and I'm excited to continue on this path.

Subscribe now

In my first YouTube video, I cover the topic of risk mitigation while holding cryptocurrency assets. I shortly introduce two essential techniques: Borrowing an asset and Hedging by opening an opposite trade in the Perpetual market.

Key Highlights:

• Providing real-life examples with screen captures for practical understanding

• Exploring market ups and downs on hedging and loans

• Explaining LTV (Loan to Value) in loans

• Understanding the Funding Rate in Perpetual markets

Check it out here:

Why Funding Rates Matter

In perpetual futures, a type of futures contract without an expiry date, the funding rate helps align the prices of perpetual contracts with the underlying asset's spot price. This mechanism involves periodic payments between buyers (long positions) and sellers (short positions), depending on market conditions. When the price of the perpetual contract is higher than the spot price, the buyers pay the sellers. Conversely, when the contract's price is lower, the sellers pay the buyers.

This topic has become increasingly relevant to me for two reasons:

1. Hedging with Derivatives: Derivatives are excellent tools for hedging. I use them to maintain stability in my portfolio, ensuring it includes only the assets I wish to hold. For instance, I recently needed USDT but only had ETH. The solution was straightforward: sell ETH on the spot market for USDT and open a Long position on Perpetuals, equivalent to the sold ETH amount. This strategy allows me to keep my ETH exposure while freeing up USDT for trading. However, it's important to note that this isn't a cost-free operation due to associated fees and the Funding Rate.

2. Profit from Funding Rate: Upon investigating the history of Funding Rates, I noticed that they are predominantly positive for most coins. So, I can use the opposite scenario and profit from Short trades. If I have extra USDT, I can, for example, allocate 70% to buy a coin like ETH on the spot market and use the remaining 30% to open a Short position in the same ETH amount. This approach helps me preserve my USDT and earn profits from the Funding Rate. Additionally, investing in a staking coin can yield extra profits. This method also can be viewed as creating a customized stablecoin with minimal or no exposure to USDT.

It's important to acknowledge the risks associated with the strategies described above

In this post, I'm focusing on studying Funding Rates. My goal is to understand the potential costs in the first scenario and the possible profits in the second. I found a useful tool on the Coinglass website that shows Accumulated Funding Rates on different exchanges. However, it only provides data for up to one year. This limitation led me to conduct my own research to uncover trends from previous years, covering both bull and bear markets.

Results

I developed a script to extract the funding rates from various exchanges, and then I grouped this data by year to calculate the accumulated totals. The results were intriguing, though most of the historical data starts from 2021. Here, I reveal my findings for the most relevant BTC and ETH markets:

The data clearly shows that during 2020-2021, the Funding Rate was high and volatile, whereas it stabilized and remained relatively low in 2022-2023. I also created a plot that correlates asset prices with accumulated funding rates on a monthly basis, illustrating that periods of high Funding Rates are closely linked with price change.

Key Insights

Two insights emerge from this analysis:

• Using 2nd strategy described above can yield profits during a crypto bull run even for USDT holdings, though this strategy requires careful handling, preferably not manual, due to holding a short position in a rising market.

• Delivery Futures might be a better hedging option since they don't involve Funding Rates. If you enter a long-term position in Delivery Futures you'll pay a fixed percentage that matches the price deviation from the spot at the time of entry. Currently, on Binance, this rate ranges from approximately 8.5 to 9.5% per year for contracts with durations of 3 and 6 months.

Of course, the data received doesn’t guarantee a repetition of past patterns in the future, and I am very curious to see how Funding Rates will behave in the coming years.

Resources

Thank you for reading. If you're interested in researching other coins or want all the data for your analysis, you can follow the GitHub link provided below.

• GitHub Historical Funding Rates Fetcher Repository

• Coinglass Accumulated Funding Rates

• How to build a stablecoin protocol from funding rate arbitrage

  Thanks for reading Viktoria's Crypto Trading Lab! Subscribe for free to receive new posts

Kucoin

Initially, I tested Kucoin using various AWS instances from around the world, employing a simple Python script to send numerous requests to the exchange endpoint. The Asia region emerged as the top performer, consistent with the results for Binance.

Further, I narrowed down the Asia region tests to four regions: Tokyo, Osaka, Seoul, and Singapore. For the Asian regions, I conducted a comprehensive test using the same methodology I applied to Binance. The specifics of this methodology can be found in my previous article.

The results for order creation are the following:

Tokyo stands out as the clear leader, demonstrating unmatched performance.

Additional Kucoin Insights

I found that Kucoin offers high-frequency pro-trade accounts. They claim that trading with these accounts results in lower latency and a looser frequency limit. I was curious about the performance of these accounts. For about 3 days, I executed real-time trading tests across various pairs on a high-activity server, setting the tick size at 0.1 sec. Here is the detailed statistics for order creation round-trip times

The data makes it clear that HFT accounts outperform the ordinary trading account. The discrepancy in Median/Average/Min/Max numbers emphasizes the superiority of Kucoin's HFT accounts over regular trading ones.

While these accounts are available to all, they can only be activated and operated via API. Additionally, certain HTTP requests, parameters, and responses for these accounts differ, requiring adjustments to the default Hummingbot settings or other bots you use. For detailed instructions, refer to the official documentation. I'll also provide a handy tool at the end of this article to help you get started with these accounts.

HTX

HTX API documentation recommends placing servers in the Japan region. A brief test on my part affirmed this recommendation.

Furthermore, HTX has an optimized endpoint specifically for AWS clients. However, upon swithching to this endpoint for both REST and Websocket, I didn't observe any significant improvements.

The outcomes were nearly identical. It's possible that a script with a higher load or testing during periods of high volatility might yield different results.

Gate.io

Gate.io explicitly states in their documentation that their servers reside in the AWS Japan ap-northeast-1 region. To verify this claim, I ran some tests for Gate.io, and unsurprisingly, Tokyo came out on top yet again:

Conclusion

Through comprehensive latency testing across several cryptocurrency exchanges, including Kucoin, HTX, and Gate.io, the Tokyo region consistently stood out as the best-performing location. Specifically, for Kucoin, not only did Tokyo showcase superior performance, but the introduction of their high-frequency pro-trade accounts further enhanced trading efficiency. Tests indicated that these accounts significantly outperformed standard trading accounts, underscoring their potential for traders aiming for reduced latency.

Futher experiments

I also experimented with AWS subnets by running multiple bots across three availability zones in the Tokyo region. Although the initial tests didn't reveal significant latency improvements for the exchanges, I see potential in this area. Conducting a more thorough analysis with an advanced script, or perhaps testing during periods of high market volatility, might produce different results.

Resources

Hummingbot - The open source OS for crypto algo traders

Crypto Exchanges Latencies Test Repository - contains the script and all data discussed in this post.

Python Script for Kucoin HFT Account Activation - utilize this script to activate the HFT account by transfering assets to trade_hf account

I started with Binance, but I have plans to study other platforms soon.

Methodology

I used Hummingbot to run the tests. As an all-encompassing open-source solution, it's equipped with multiple exchange connectors. It's also user-friendly for creating crypto trading bots, achievable with just a single Python script.

The primary goal wasn't to find exact latencies, but rather to determine the best AWS region. To achieve this, I devised a straightforward hummingbot script to compare the time durations for order creation, cancellation, and execution.

Here's how it works:

1. Timestamps: For every order – whether created, cancelled, or executed – the script records two timestamps along with the order ID in a CSV file.

   • Pre-transmission Timestamp: This marks the time just before the API request is sent to the exchange.

   • Post-transmission Timestamp: This is the time when a successful confirmation is received from the exchange.

   By analyzing the difference between the pre-transmission and post-transmission timestamps, one can determine the round-trip time for each order activity.

2. Order Types and Frequencies:

   • Limit Orders: I tested creation and cancellation by generating limit orders every 30 seconds.

   • Market Orders: Execution was assessed using market orders created every 5 minutes.

3. Websocket and REST API: While actions like order creation, cancellation, and execution are transmitted using the API request, confirmations are retrieved via Hummingbot events, which are dependent on a Websocket connection.

AWS Setup

I set up identical t3.small instances across various regions, all using the same image for consistency. The only distinction between them was the connection to individual Binance subaccounts. This was done to ensure that there were no mix-ups or overlaps in API requests.

Results

My testing consisted of two stages: I started with a quick high-level test across continents and then went deeper into the best-performing area.

The first test took place across 8 regions within a 4-hour window:

• Asia-Pacific: Tokyo, Singapore

• Europe: London, Frankfurt

• North America: Canada

• South America: São Paulo

• Middle East: UAE

• Africa: Cape Town

Below are the box plot results specific to order creation:

From the results, it's evident that the Asia region outperforms the rest.

Building on that, I decided to delve deeper into Asia, launching servers in seven distinct regions:

Tokyo, Osaka, Seoul, Hong Kong, Singapore, Mumbai, Sydney

This time, the bots operated for a full 24 hours. Each bot processed about 4000 limit orders—both creating and canceling—and successfully executed around 200 market orders.

Below are the box plots illustrating order creation, cancellation, and execution:

Accompanying that is the statistical tables:

Order creation round-trip latencies statistics for Asia region:

Order cancellation round-trip latencies statistics for Asia region:

Order execution round-trip latencies statistics for Asia region:

To my surprise, the winner was not Tokyo but Osaka. In general, Seoul sometimes showed better results than Tokyo as well. I decided to run another test, focusing solely on these three regions for an additional 24 hours. The results were consistent: Osaka performed better, while Seoul and Tokyo competed closely for second place.

Conclusion

In conclusion, this analysis elucidates the significant impact of server location on trading bot latency on Binance across various AWS regions. The findings notably highlight the superior performance of servers in the Asia region, particularly in Osaka, Tokyo and Seoul, in terms of order creation, execution and cancellation latencies. This insight is crucial for crypto traders looking to optimize their trading bots for faster and more efficient order processing. As the crypto trading landscape evolves, understanding such nuances in server latency across different regions remains imperative for gaining a competitive edge in the market.

What's Next?

Next, I plan to conduct similar tests for other exchanges. I'd love to know which ones you'd like me to explore – drop your suggestions in the comments! Additionally, any insights to enhance the methodology would be greatly appreciated. I'm always open to feedback, constructive criticism, and advice. If you fancy a chat or even a quick call, don't hesitate to reach out to me on Discord

Resources

Hummingbot - The open source OS for crypto algo traders

Crypto Exchanges Latencies Test Repository - contains the script and all data discussed in this post.