Methodology
Anyone can produce a beautiful equity curve. The skill is knowing which ones are real — and we apply all four tests to ourselves, in public.
Every decision uses only information available before the bar it trades. Decide on the closed candle; fill the next one. It's the most common way backtests lie.
An early engine showed MNQ +$22,000. The signal was peeking at the same bar's close. We fixed the clock — strictly causal — and the same strategy printed −$13,000. That ~$35k swing was the look-ahead. A profit that dies when you fix the clock was never profit.
We evolve on the past, then test on years the optimizer never saw — rolling forward, block by block. Stitched together, that's a true out-of-sample record, not a fitted one. Our intermarket scorer was positive in 6 of 6 unseen blocks.
If you try 6,000 strategies and keep the best, the "best" looks amazing by luck alone. The Deflated Sharpe (Bailey & López de Prado) corrects for exactly that.
Flip 3 coins; ~12% of the time you get 3 heads. Flip 200 sets; ~25 look "perfect." Pick the best and it looks like genius — it's noise. Deflated Sharpe puts a number on that.
Our most-evolved router had a low deflated Sharpe — so we don't trade it. We deploy the simpler exposure that doesn't depend on selection. Most shops bury this number. We lead with it.
Commissions, spread, slippage and gap-aware stop fills are in every test — stops can fill worse than their level on a gap, never magically at it. A strategy that only works at zero cost is a fee generator, not a strategy.
The tell to use on any bot: ask for the deflated Sharpe and the walk-forward. A great backtest Sharpe with no deflated number is a marketing asset, not a track record.