Three timeframes, 231 rules, and no edge worth trading

The result does not depend on which cost you believe

A rule only matters if it beats holding the stock. Ranking on raw Sharpe would mostly rank how long-biased each rule is, which in a rising market flatters anything that stays invested — so every figure here is excess over buy-and-hold on the same ticker, over the same bars. Zero is parity. Above zero is an edge.

The chart below tracks the single best rule at each cost level. If technical signal existed and costs were merely eating it, the lines would start above zero and fall through it. Instead they start at or below zero and go down — the signal is not there to be eaten.

The leaderboards, in full

Ranked by average excess Sharpe across the twenty stocks at per unit of position change. Bars run left from the zero line, so a longer bar is a worse rule. Beat rate is the share of the twenty stocks where the rule actually finished ahead of buy-and-hold.

What this means for the machine-learning step

The useful part of a null result is that it narrows the search. These rules failed in a specific way, and that shape is what the next stage should be built around.

• don'tFeed a model these 231 signals and ask it to pick winners. There is nothing to pick — the best rule at every timeframe and every cost level is at or below parity, so a selector trained on this space learns noise.
• doTreat the indicators as features, not strategies. A rule that is flat on average can still carry conditional information; that is a different question than the one tested here, and it is the one worth asking a model.
• doPredict something other than direction. Volatility, drawdown risk and position sizing are more learnable than sign, and they are where a model can add value on top of a buy-and-hold core.
• watchTurnover is the binding constraint. Any candidate that trades more than a few dozen times a year has to clear a cost bar that killed everything on this page — check it before, not after, the modelling.