blog.
Backtesting, scanning, auto-trading. Notes on treating trading as data, not gut feel.
FEATUREDBacktesting
The Random-Max Percentile: Grading Every Backtest Against Its Search's Random Maximum
RMP is a new column in our scan results: the probability that the best result pure chance could produce, across the N trials your scan actually ran, lands below this row. We built it because deflated Sharpe goes quiet at high trade counts: in one 5.81-million-trial search from our audit, the best DSR at 3,000 trades or more reads 0.000, real edge or not. RMP tracks PSR's ordering, carries a search-sized hurdle, and stays readable exactly there.
Latest posts
2026.07.22What Should You Change First in a Crypto Backtest? 99.75 Million TestsWe crossed 79,800 entry rules with 125 exit rules across ten crypto market-direction jobs to measure which settings actually move return, drawdown, win rate, trade count, and an out-of-sample proxy. There was no universal winner: the right tuning order depends on the result you want to change.2026.07.20We Ran 99.7 Million Crypto Backtests. One Market Produced All 103 Candidates.We scanned BTC, ETH, BNB, XRP, and SOL in both directions. A post-analysis RMP screen of 99,691,200 results left 103 research candidates, all in one market and direction.2026.07.13Your Backtest Dies Climbing Into MajorsWe ran the same 48-million-backtest search on three tiers of coins, majors down to thin small-caps, then looked up every strategy that beat one tier's luck bar inside the other tiers' results. Three layers of statistics later, one number was left standing: transplants into BTC, ETH, SOL and BNB passed a fair test zero times out of 379 checks. Chance alone should have passed about 103.2026.07.10How Many Backtests Is Too Many? We Ran the Same Search Twice to Find OutWe ran one 5.7-year crypto search twice, removing four rarely-firing indicator slots the second time. That one change cut the luck bar for every other strategy by more than half, and we predicted the second run's numbers from the first run's export before it started. Here is the dial that actually moves your overfitting hurdle.2026.07.08Your AI's Best Backtest Can't Compute Its Own Deflated SharpeLearn how much of your AI's best backtest is search luck. A lone backtest grades itself against zero; a scan grades it against the 893 variants it beat to look that good. Watch a Sharpe 4.514 short deflate to a Deflated Sharpe of 0.000, then download every candidate with its own dsr and judge the rest yourself.2026.07.04We Combined 1.1 Million Strategy Pairs. Three in Four Diluted the Better One.We combined all 1,131,321 pairs of our gauntlet-surviving Bitcoin strategies with zero extra backtests. Only 25% beat simply using the better strategy alone, yet 92% still cleared the luck baseline. Here is how to tell when combining two trading strategies helps and when it just dilutes your best idea.2026.07.03We Ran 100 Million Bitcoin Backtests. Only 3,130 Beat the Luck Bar.We ran 99,878,688 Bitcoin backtests and measured each with the Deflated Sharpe Ratio. Only 3,130 (0.003%) beat the luck you would expect from searching that hard, and the only result to clear the strictest bar was two survivors pooled into one. Here is the step-by-step method.2026.06.13Your Crypto Backtest Is Lying to You: 3 Reasons WhyA backtest that looks great can still lose money live. The cause is usually not the strategy but how it was measured. Here are the three measurement flaws, and how to catch each.2026.06.13How to Read Your Backtest Results: Sharpe, Profit Factor, Max Drawdown, Win RateWhat each backtest metric measures, how to read it, and where it fools you. Read Sharpe, profit factor, max drawdown, and win rate as a stack, not as one number.2026.06.13Repainting: Why a Trading Signal Should Only Count After the Candle ClosesRepainting is when a signal that appeared mid-candle vanishes once the candle closes. It is look-ahead bias on a live chart. Here is how to spot it and prevent it.
