424,471 captured 5-minute windows, 5 imbalance buckets, and a permutation null baseline. No reliable edge found: the observed bucket spread is consistent with random reassignment (permutation p-value 0.6577) on 424,471 captured windows.
| Bucket | Imbalance range | Windows | Mean forward return (bps) |
|---|---|---|---|
| 1 | -0.98 to -0.21 | 84,895 | +0.37 |
| 2 | -0.21 to -0.06 | 84,894 | +0.25 |
| 3 | -0.06 to +0.06 | 84,894 | +0.54 |
| 4 | +0.06 to +0.20 | 84,894 | +0.54 |
| 5 | +0.20 to +0.98 | 84,894 | +0.12 |
Result. No reliable edge found: the observed bucket spread is consistent with random reassignment (permutation p-value 0.6577) on 424,471 captured windows.
From our own capture of public exchange feeds, each 5-minute window contributes a taker imbalance (aggressive buy volume minus aggressive sell volume, scaled to the range -1 to +1) and the forward 5-minute return. Windows are grouped into imbalance buckets, and equal imbalance values always share a bucket — a tie is never split across buckets — and the mean forward return is measured in each bucket. The null baseline asks whether the observed spread between bucket means is larger than random reassignment of the same returns would produce. Because symbols move together, windows captured at the same moment are reassigned together as one block, so the market-wide move they share stays intact in the baseline: the blocks are reshuffled many times with a fixed random seed and we report the share of reshuffles that match or exceed the observed statistic (a permutation p-value). A p-value below 0.05 is reported as a candidate edge; otherwise the study says plainly that no reliable edge was found.
This measures historical association, not tradeable alpha: it excludes fees, funding, slippage and market impact, and it does not adjust for looking at many symbols and thresholds. A low p-value can arise by chance when many tests are run, and an association measured on captured windows is not evidence of a profitable strategy. This is educational research, not a signal.