Does any TA-Lib rule beat buy-and-hold across crypto and US stocks?

Methodology

Why this is ranked by information ratio, not by profit

Every rule here is scored on information ratio against buy-and-hold on the same asset — not on dollars earned, not on raw Sharpe. Those two reward something that looks like skill and is not.

IR = mean(r − r_buy&hold) ÷ std(r − r_buy&hold) × √(periods per year)

Read it as average lead ÷ how much that lead wobbles. Two rules can finish equally far ahead; the one that got there steadily is skilled, the one that got there in a single lucky year is not. Dividing by the wobble separates them. Doing nothing different from the benchmark scores exactly zero.

What profit cannot tell you

• It rewards exposure, not accuracy. In a rising market a rule that is simply long more often earns more, with no predictive ability whatsoever. Dollars cannot distinguish was in the market from was right.
• It is dominated by the largest winner. Pooled dollars across a universe are carried by whichever asset trended hardest, drowning out the verdict on every other name. IR averages per asset, so each one gets an equal vote — which is what makes the breadth gate possible at all.
• It has no risk term. Same profit, wildly different paths, identical ranking.
• It has no benchmark. A large number sounds good until you learn what holding would have paid for free. IR bakes the comparison in: negative means it lost.
• It is not comparable across timeframes or asset classes. IR is annualised and unit-free, so the same value means the same thing everywhere.

This project has already been caught by exactly that trap: an earlier sweep in the same repo produced “winners” that turned out to be long-bias in a rising market, because it ranked on dollar PnL and raw Sharpe. The scoreboard cards still display dollars, but nothing is ordered by them.

The four gates

Ranking is not the same as passing. The leaderboard orders candidates by IR, but a candidate only counts as an edge if it clears all four gates at once — out of sample, at the real fee schedule. Each gate kills a different way of being fooled, which is why one strong number is never enough.

Notice that each gate has a “too good” band. A result can fail by being too strong: an IR above 2, perfect breadth, or a t above 6 on a sample this short is far more likely to be a look-ahead leak, an accounting error or a stale-data artifact than a discovery. Every one of those has happened in this repo. Treat a sudden winner as a bug until the parity harness and the multiplicity correction have both been re-checked.

Why the gates are this hard

The governing constraint is t = IR × √years, and it is unforgiving. On an 11-year sample √11 ≈ 3.3, so even a genuinely good IR of 0.5 reaches only t ≈ 1.7 — below the bar. A system can be real and still not provable on the history that exists.

Two consequences follow, and neither is optional:

• A finer timeframe buys no significance. t scales with the square root of the sample length, not the bar count. Running the same six years at one minute instead of daily gives 390× the bars and an identical √years.
• Searching harder raises the bar rather than lowering it. With hundreds of candidates the best one is an extreme-value statistic, so the noise ceiling — the IR the best of N worthless rules reaches by luck — is quoted beside every leaderboard maximum, along with the Bonferroni-corrected t threshold that candidate count implies.

Only two levers actually move these gates, and neither is “try more indicators”: more history (t scales with √years) and lower turnover (fee headroom scales inversely with it).