Why I built StratProof

The short version

A couple of years ago I built a trading strategy that looked great in a backtest. Equity curve climbing left to right, sharpe over 2, drawdown manageable. I funded it with real money. Over the next several months I gave most of that money back to the market.

The backtest wasn't wrong on purpose. It just assumed zero slippage, same-candle fills, no funding costs, and the same 18 months I'd overfit my parameters to. When I re-ran it with realistic fees and a different time period, the edge disappeared. I hadn't discovered an edge. I'd discovered a way to make overfitting look like one.

StratProof exists so anyone testing an idea gets told the truth the first time, not the fifth. You type the strategy in plain English. I run it through 3 years of Binance data on 10 coins across 4 timeframes, with real exchange fees, next-candle entries, and per-coin spread models. You find out in under two minutes whether the idea survives contact with a realistic market or dies the moment fees are included.

Most strategies die. Mine did. That's fine. Knowing before you fund it is the whole point.

Who runs it

Solo founder. Software engineer by background. I've been trading crypto for about five years, mostly losing, finally learning. I wrote every line of this codebase, I run the VPS, and if you email support@stratproof.com it hits my inbox directly.

No team, no VCs, no paid advisors pretending to be objective. The site pays for itself through a small founding-member tier, and the research engine is a mix of curiosity and a way to validate my own trading ideas before I risk capital on them.

How the engine tests your strategy

Every strategy runs through the same pipeline, with no knobs or exceptions for anyone. Same treatment whether you're testing RSI on Bitcoin or a grid-bot proxy on a memecoin.

• Real exchange fees.Binance maker/taker rates baked into every trade. If you'd pay 0.075% per side with BNB discount, that's what the backtest pays. Strategies that only work at zero fees are flagged as such.
• Realistic slippage. 0.1% adverse fill on every entry and exit, calibrated against our own live paper trades. Not the zero-slippage fantasy most backtesters default to.
• Next-candle entry.You can't trade on a candle that hasn't closed yet. When your signal fires, we enter at the next bar's open. Other backtesters enter on the same candle and inflate results.
• Per-coin spread model. 14,000+ L2 order book samples across 38 coins. Thin books cost more than thick ones, and the backtest knows the difference.
• Walk-forward validation.The strategy gets optimized on 60% of the data it can see and tested on the 40% it can't. If performance collapses on the unseen data, the test is marked as overfit.
• Multi-coin, multi-timeframe.Every strategy runs on 10 coins and 4 timeframes in parallel. If it only works on one coin or one timeframe, that's called out. Single-coin edges are usually coincidence.
• 3 years of data. Covers the 2023 bull, the 2024 grind, and every regime shift in between. Not a cherry-picked 6-month window.
• Fixed 2x/4x ATR exit structure. Every entry gets the same stop (2x ATR below) and take-profit (4x ATR above). This is a 2:1 reward-to-risk ratio, which puts the mathematical breakeven win rate at 33%. If you run a strategy and see a 28-32% win rate, that means the signal is landing close to random-walk expectation. To clear our costs, a strategy needs either a real directional edge (pushes win rate above 40%) or a structurally different exit, which the research lab uses and the free tool intentionally does not. We keep the free tool exit fixed so results are apples-to-apples across every strategy anyone tests.

What the test doesn't capture

Honesty goes both ways. Here's what the engine doesn't know:

• • 3 years of data is a lot, but a once-in-a-decade event could change the answer for any strategy.
• • We test Binance spot and linear perps. Perp funding rates aren't factored in yet. Usually a small tailwind, not always.
• • Real fills have edge cases our model approximates: partial fills during illiquid moments, order book depth under stress, latency during fast moves.
• • If the AI parser misinterpreted your plain-English description, the test is for a different strategy than you meant. You can always check and correct the conditions on the report page.
• • We don't model trailing stops, DCA ladders, grid position management, or funding-rate carry strategies as full systems. We test their entry signal as a proxy and mark the difference.

The research lab

There's a second engine running behind the scenes, independent of the public Prove It flow. It continuously tests candidate strategies against the same engine, with the same calibration, looking for real edges that survive walk-forward validation and forward testing with live paper money.

Nothing makes it out of the lab just because the in-sample backtest looked good. A candidate has to pass five statistical gates in backtest, then accumulate at least 8 forward-test trades with a 90% confidence lower bound above zero, before it shows up anywhere as a "ready" strategy.

The audit trail for every candidate, what it tested, what survived, what got killed and why, lives at /architecture.

What I commit to

• • I will never hide a losing result to make a strategy look better.
• • I will never publish a number I can't reproduce from the database.
• • I will never sell you a strategy with a claimed track record. This is a testing tool, not a signal provider.
• • If I change the methodology in a way that affects past results, I will note it publicly and flag affected reports.
• • If you email me about a bug or bad result, I will read it and respond personally.