VERDICT 0003 · RSI mean reversion ("oversold = buy")
BTC + ETH · 4h · walk-forward · 32-combo sweep
BEST IN TRAINING
+307.8 bps / trade
SAME COMBO, VALIDATE
−45.6 bps
SIGNIFICANCE
t = −0.1
(needs t > 2 to beat luck)
PROTOCOL
commit to best-on-train combo, validate once, t-test vs noise
BELOW BAR — DO NOT TRADE
The most-believed setup in retail trading. In training, the best RSI threshold returned +307 bps per trade — a money printer. Committed to that exact setting, tested once on unseen data: −45.6 bps, with a t-stat of −0.1. Every other "profitable" setting in the sweep was statistically indistinguishable from a coin flip. The edge was never there. It was the search finding the prettiest accident.
VERDICT 0001 · order-flow momentum signal
BTC·ETH·SOL·XRP·AVAX · L2 tape · walk-forward
TRAIN edge
+10.3 bps
VALIDATE edge
−8.3 bps
COST bar
6.0 bps round trip
PROTOCOL
120-combo sweep · train first half · one-shot validate · 8 finalists, all below bar
BELOW BAR — DO NOT TRADE
An order-flow momentum signal. Every one of the 8 best training combos was profitable in-sample (+8 to +10 bps) and every one inverted to a loss on unseen data (−7 to −9 bps). The edge didn't fade to noise — it flipped negative. Nothing cleared the cost bar out of sample.
VERDICT 0002 · ICT v4 perpetuals strategy
crypto perps · 2024–2026 · 298 trades
DURING TUNING
+233.1%
FINAL CANDIDATE
−31.2%
FINAL WIN RATE
33.2%
(99W / 199L) · 0.78× PF · Sharpe −1.59
DRAWDOWN
35.2% · 13 losing months
BELOW BAR — DO NOT TRADE
Earlier versions backtested +224% and +233%. So we kept "improving" it — raising thresholds, adding a daily-bias multiplier, chasing the upward trend. The final optimized candidate lost −31.2%. The improvement was the overfitting tightening: every tweak fit the past more perfectly until the strategy had memorized noise. The bot was built and deployed before this was understood — and shut down before a customer ever paid for it.