# mean_reversion / 3

Universe-generic version: META["universe"] = ["*"]. The symbols are not written in the strategy; they are the ones the agent trades (1 to 12 market symbols), passed by the engine with the candles and by --symbols to the backtest. Same rule and same parameters as version 1: nothing was re-tuned, for any universe. Each symbol is judged against its own recent average; N is the number of symbols the agent chose.

What a generic universe changes. The code hash still pins the behaviour, and the agent's mandate pins the assets. But the evidence does not travel: a backtest report is attached to ONE agent and records the symbols it was run on (config.symbols); a result on one universe says nothing about another, and every universe tried is one more trial. Symbols with a short history shorten the common backtest period for all of them.

Hypothesis. Over a couple of days, a sharp move away from the recent average price is partly a liquidity shock that reverts: an asset trading well below its 48-hour mean tends to bounce back towards it (the lower Bollinger band as a buy zone).

Rule. z = (close - mean48) / std48. Weight per symbol = (1/N) * clip(-z / z_full, 0, 1): nothing at or above the mean, scaling in linearly below it, fully in at z <= -2. There is no memory: the position shrinks as the price returns to its mean. Long-only, so it never sells the upper band short.

What would refute it. It buys falling prices: in a sustained down-trend it is repeatedly long while the price keeps falling ("catching a falling knife"). It trades often, so costs weigh heavily; with 20 bps per side the average bounce must be several times larger than that to leave anything.