Chaikin Money Flow

On data it never saw, the agent averaged -5.4bp per trade after costs (win rate 33.3% — irrelevant when losers outweigh winners). Historically, 37.1% of ALL graded strategies also fail this (N=143).

Try: This is an entry-signal problem, not a tuning problem. Change the core idea: different condition, different timeframe, or add a regime filter (e.g. only trade high-volatility periods) — don't just widen stops.

It loses on AVAX/USDT:USDT, BTC/USDT:USDT, LINK/USDT:USDT, SOL/USDT:USDT, XRP/USDT:USDT — under 60% of its universe is positive, which usually means the backtest found luck, not a mechanism. Historically, 37.8% of ALL graded strategies also fail this (N=143).

Try: Either the thesis only applies to certain assets (then say so — trim the universe to where the mechanism makes sense), or the signal is noise. Ask: WHY would this work on some symbols and not others?

Simulated at responsible sizing, the worst drawdown hits 81% — deep enough that most people abandon the strategy exactly at the bottom. Historically, 25.9% of ALL graded strategies also fail this (N=143).

Try: Cut risk per trade, tighten the catastrophe stop, or reduce concurrent positions. An edge you can't hold through its drawdown is an edge you don't have.

Performance collapses between the training and test periods — the parameters fit historical noise, not a repeatable mechanism. Historically, 4.2% of ALL graded strategies also fail this (N=143).

Try: Simplify: fewer conditions, rounder parameter values (RSI 30, not 28.7), wider regime windows. If a small parameter change kills it, it was never real.