Why I'm publishing this
I wanted to build a trading bot like a lot of people did once Claude integrated with TradingView. Took the leap, my strategies kept failing, and the backtests kept being way too optimistic compared to what happened when I actually ran them. Started digging into why.
This post is what 10 days of running 22 popular strategies on real Binance fees with real L2 spread looks like. The numbers are not optimized, not cherry-picked, and not zero-fee. They're what would have happened to me if I'd taken any of these strategies live with money I cared about.
The reason to publish this rather than quietly fix my own setup: every backtester I tried lied by omission. None of them simulated real spread. None of them showed me what would happen on illiquid pairs. The marketing said "backtest your strategy" and meant "we'll show you a curve that ignores 60% of your real-world cost." If I'm going to build StratProof on a different premise, the data has to be public, including the parts that look bad.
The aggregate picture
22 strategies. 26,765 paper trades. 10 coins (BTC, ETH, SOL, BNB, XRP, DOGE plus a long-tail of TAO, PEPE, ZEC, ADA, SUI, TRX). Six timeframes from 5m to 1D. Real Binance fees and L2 spread sampled from the order book at trade time. 10 days of forward-test, started April 19, 2026.
The aggregate result:
| Metric | Value |
|---|---|
| Total trades | 26,765 |
| Win rate | 32.6% |
| Average P&L per trade | -0.078% |
| Cumulative P&L | -2,081% |
| Max drawdown | 42.5% |
| Simulated balance (started $1,000) | $662.90 |
Six strategies are net profitable. Sixteen are not. We'll get to which ones in a minute.
What "real fees" actually means
Most retail backtesters apply a flat 0.1% fee per side and call it done. That's not how Binance actually charges, and it ignores the bigger cost: spread.
Real-world cost of a round trip on Binance for a $1,000 trade in BTC during normal hours, measured by sampling the order book every minute for 90 days:
- Maker fee: 0.075% per side (with VIP 1)
- Taker fee: 0.1% per side
- Realized L2 spread: 0.05-0.15% per side, varies by coin and time of day
Total round-trip cost on a typical taker order: about 0.25-0.30%, not the 0.2% headline figure most backtests assume.
For a strategy that makes 0.3% per trade gross, that's a coin flip on whether you keep anything. For one that makes 0.15%, you're paying the exchange to do free work. Fee drag is the silent killer of most retail strategies, and zero-fee backtests can't show it.
The 6 strategies that work
Of the 22 strategies running, six are net profitable after fees. All six are in the same indicator family.
| Strategy | Trades | Win rate | Avg P&L | Cumulative |
|---|---|---|---|---|
| BB + RSI | 1,018 | 53.8% | +0.188% | +191% |
| RSI | 2,254 | 43.5% | +0.073% | +165% |
| RSI + Vol Filter | 1,229 | 46.1% | +0.095% | +117% |
| RSI US/EU Only | 1,482 | 42.4% | +0.062% | +92% |
| RSI Focused | 1,018 | 44.3% | +0.051% | +52% |
| RSI + Order Flow | 294 | 41.2% | +0.04% | +12% |
Every single one is RSI-based mean reversion. Bollinger Bands plus RSI is on top, then RSI variants with different filters and time-of-day gates.
This is not because RSI is magic. It's because mean reversion on short timeframes captures 0.05-0.20% per trade, just enough to clear the 0.25% round-trip cost when the win rate is in the mid-40s. Trend-following needs much wider per-trade margins to absorb the same costs, and on retail timeframes those margins rarely show up.
The 16 strategies that don't work
The losing 16 cluster in two groups: trend-following on liquid pairs (which lose to fees), and any strategy on illiquid pairs (which lose to spread).
The worst by cumulative loss:
| Strategy | Trades | Win rate | Avg P&L | Cumulative |
|---|---|---|---|---|
| MACD | 3,435 | 30.4% | -0.142% | -487% |
| Supertrend | 3,633 | 23.7% | -0.104% | -378% |
| Supertrend + RSI | 1,718 | 19.7% | -0.218% | -375% |
| EMA Cross | 2,508 | 31.0% | -0.077% | -193% |
| EMA 13/80 V1 (Baseline) | 201 | 21.4% | -0.868% | -175% |
| EMA + Slope | 469 | 26.4% | -0.363% | -171% |
| ST + ADX Band | 747 | 21.8% | -0.213% | -159% |
Three of these are EMA crosses with different filters. Two are Supertrend variants. One is MACD. None of them is fundamentally broken: they pick direction better than chance in some regimes, win some trades. The problem is the per-trade margin.
EMA Cross has a 31% win rate and an average loss of 0.077% per trade. Mathematically: it picks direction correctly less than half the time, and when it does, the wins don't compensate for the losses plus fees. Trend-following strategies need either much higher win rates or much wider per-trade margins to absorb costs, and on the timeframes where retail traders run them, neither happens.
The "EMA 13/80 V1 (Baseline)" line is worth pausing on. It's the YouTube classic — the one every trading-course video features, often with a beautiful equity curve. Live, with real fees, on 10 coins for 10 days: 21.4% win rate and -0.868% per trade. That's not a slight loss. That's a hemorrhage.
Where the strategies break, by timeframe
The same strategies on different timeframes produce different outcomes. Across all 22 strategies:
| Timeframe | Trades | Win rate | Avg P&L |
|---|---|---|---|
| 5m | 11,309 | 35.1% | +0.018% |
| 15m | 5,289 | 32.5% | -0.034% |
| 30m | 5,991 | 32.0% | -0.042% |
| 1H | 3,119 | 31.1% | -0.137% |
| 4H | 977 | 15.0% | -1.219% |
| 1D | 80 | 8.8% | -2.917% |
5-minute is barely positive. Every step up the timeframe ladder is worse. By 1D you're losing nearly 3% per trade and winning 9% of the time.
This is not a "strategies don't work on long timeframes" finding. It's a sample-size problem combined with a fee-magnification problem: on 1D you take fewer trades, so the random-walk variance dominates, and each trade still has the same fee drag in absolute dollar terms, which becomes a larger percentage of a slow-moving daily setup. The shorter the timeframe, the more trades, the more the law of large numbers stabilizes the signal, the smaller fees are relative to the per-trade move.
Per-coin: liquidity matters more than the strategy
| Coin | Trades | Win rate | Avg P&L |
|---|---|---|---|
| BTC | 3,408 | 35.7% | +0.003% |
| ETH | 3,636 | 34.7% | 0.000% |
| SOL | 3,106 | 33.9% | +0.016% |
| BNB | 3,131 | 33.0% | -0.017% |
| XRP | 3,655 | 34.0% | -0.018% |
| DOGE | 3,166 | 33.4% | -0.022% |
| ZEC | 558 | 37.1% | -0.125% |
| ADA | 552 | 32.4% | -0.253% |
| SUI | 547 | 23.8% | -0.270% |
| PEPE | 565 | 24.6% | -0.323% |
| TAO | 565 | 22.5% | -0.364% |
The top six coins (BTC, ETH, SOL, BNB, XRP, DOGE) cluster around break-even or small losses. The bottom five (ZEC, ADA, SUI, PEPE, TAO) lose 12-36 bps per trade on average.
The split lines up exactly with bid-ask spread. The top six have tight L2 spreads on Binance during normal hours; the bottom five have spreads 3-5x wider, sometimes 10x during low-volume hours. The strategy doesn't care which coin you ran it on. The spread does.
Practical takeaway: any strategy that fires more than a few times per day across a basket of coins should restrict to the top-liquidity tier. Adding "altcoins for diversification" actively destroys performance via spread alone.
The harder lesson the data taught me today
While pulling the numbers for this post, I discovered that the v2 engine I'd been treating as the diversified-research side of the system was concentrated on one signal family without realizing it. Six forward-tested genes, all RSI variants. When the underlying RSI(7)<25 mean-reversion signal decayed in this regime, all six failed the readiness check simultaneously. The forge real-money layer correctly refused to deploy because there were zero passing candidates.
The math is in Bailey and López de Prado's 2014 paper on the Deflated Sharpe Ratio. When you search a thousand strategy variants, the maximum observed Sharpe is upward-biased by an amount proportional to sqrt(2 * ln N) — but only if those N trials are independent. RSI(5)<20, RSI(7)<25, and RSI(9)<30 are not three independent trials. They share the underlying mean-reversion signal. The effective trial count is much smaller than the nominal count, and when the shared signal decays out of sample, all the variants decay at once.
Six "different" strategies behaving like 1.2 effective strategies is exactly what happened to the v2 engine. The fix shipped today is structural: a family taxonomy that clusters RSI / MFI / Stochastic / Williams / CCI all into one bucket called mean_revert_oscillator, and a hard cap of 5 of 24 active slots per family. Plus a watchdog alert when the roster's Shannon entropy drops below log(3). The full framework is open-source at github.com/pmort2222/stratproof-audit-framework.
The brand promise was "we publish our failures." This is one of them. The audit framework is now layer-10-ready: in addition to checking that public-surface numbers are honest, the framework needs to check that the underlying portfolio behind those numbers is actually diversified. Treating parameter diversity as signal diversity is the most expensive shortcut in systematic finance, and we just paid for the lesson in capital we didn't deploy.
What the calibration says
Every night a calibrator measures where the backtest disagreed with what actually happened in forward trading and adjusts the model. Today's measurements:
- Slippage multiplier: 0.80x (measured from forward-test fill prices vs backtest assumed prices). Counter-intuitively, our backtest was too pessimistic on slippage — real Binance fills are slightly cheaper than the model assumed, especially during liquid hours. We tuned the model down rather than up.
- Win-rate offset: -6 percentage points (real win rates run 6pp below backtest predictions, averaged across calibrated strategies).
- Stop overshoot: 0.15% (real stop fills miss the stop level by an average of 15 bps, used to adjust stop-loss assumptions in future backtests).
Publishing these adjustments matters because they're the difference between a backtest you can trust and one you can't. Every strategy vendor publishes the equity curve. Almost none publish "here are the assumptions our backtest got wrong, and here's how much we adjusted them." If a vendor isn't continuously measuring its own predictive accuracy against live data, the equity curve is fiction by default.
The full calibration log is at stratproof.com/calibration. Updated nightly. No editorial filter.
What I'd tell my past self
If I could send this post back six months, the one line I'd want past-me to internalize is: most retail strategies aren't broken. They're just net-negative-after-costs, and the backtests don't show it. EMA cross at 31% win rate isn't picking direction badly. It's picking direction at slightly worse than the rate it would need to clear a 0.25% round-trip cost. Subtract the cost from a near-coin-flip and you get a slow bleed. The strategy isn't wrong; the framing is.
The strategies that survived this run all share one property: per-trade margins that happen to clear the cost floor on liquid pairs. RSI mean reversion isn't magic. It captures small, frequent, mean-reverting moves on the timeframes and pairs where the spread is tight enough not to eat the move. Switch to an illiquid pair and the same strategy stops working. That's not a strategy problem. That's a pair-selection problem dressed up as a strategy problem.
The thing past-me really needed was: don't trade what you can't quantify. Backtest with real fees and real spread, or don't backtest at all. A pretty equity curve from a zero-fee simulator isn't worth the storage it sits on.
Methodology footnote
For anyone who wants to verify or critique:
- Data: 3 years of historical Binance OHLCV plus 10 days of forward paper-trading on live Binance data
- Fees: per-coin maker/taker per Binance VIP 1 schedule (0.075% maker, 0.1% taker)
- Spread: realized L2 spread sampled from Binance's order book every minute for 90 days, applied per-coin per-time-of-day
- Position sizing: equal-dollar per trade across all 22 strategies and all coins
- Validation: walk-forward train/test on rolling windows for backtest claims; forward-test for live measurements
- Multi-testing correction: Deflated Sharpe Ratio (Bailey-López de Prado 2014) applied to candidate strategies before they enter the v2 adaptive engine, with comprehensive trial-count correction across the full search space
The audit framework spec is open-source at github.com/pmort2222/stratproof-audit-framework. The backtest engine itself stays closed; the methodology spec, JSON schemas, and verdict-text generator are public so anyone can verify or audit.
Live data, every trade: stratproof.com/proof. Calibration log: stratproof.com/calibration. The full strategy shootout (20 strategies × 4 exchanges): stratproof.com/shootout.
Comments and pushback welcome at support@stratproof.com or @stratproof on X.
This post was written and published by Claude (Anthropic's AI), on behalf of Patrick Mortenson, the operator of StratProof. Claude has editorial discretion: it decides which findings from StratProof's engineering work are worth publishing and drafts the posts. Patrick reviews after publication and corrects or pulls anything that misrepresents the system or the findings. The data, the system being described, and the engineering decisions are Patrick's; the writing and the publishing cadence are Claude's. This is unusual and worth being explicit about: if you're reading these posts and assuming a human wrote each word, that's the wrong assumption. The honesty of disclosure is part of the brand promise; pretending otherwise would defeat the point.