25 Trading Strategies Tested, 0 Promoted: A Falsification Log

A falsification log from a real research program. 25 hypotheses, pre-registered. Zero survived.

Most trading content sells you the dream: the winning strategy, the secret indicator, the backtest that goes up and to the right. This is the opposite. Over the last year I ran a systematic research program testing retail algo trading strategies — every hypothesis pre-registered, every test run through identical gates (baseline → costs → walk-forward, no tuning). The result:

25 strategies tested. 0 promoted. 13 dead ends documented with exact numbers.

The scoreboard (abridged)

• Cross-sectional momentum (BTC/ETH): Sharpe 0.60 vs 0.53 buy-and-hold. Last walk-forward fold: −43.3%. Rejected.
• Volatility breakout: Sharpe 0.21 vs 0.66 benchmark, 7 trades in 2.75 years. Rejected — signal too rare to matter.
• Funding-rate contrarian: Sharpe 0.33, ~10 entries in 2.75 years. Rejected.
• Intraday BTC mean-reversion at retail fees: loses in both directions once you model 10bps fees + 5bps slippage per side. Structurally unprofitable.
• AI-assisted strategy selection vs control: 0/3 vs 0/3 promoted. No AI advantage found.

The retail fee wall

The single most important finding wasn't about any strategy — it was about arithmetic. At retail fee tiers, any edge under ~25–30bps gross is untradeable. Fees don't just reduce returns; they convert small edges into certain losses. Most published "profitable" backtests quietly assume institutional fees the reader will never get.

What the global evidence actually supports

I also ranked seven trading mechanisms by evidence grade, from B down to F: