Case 8 · Data semantics
Case 8 — A column that changed meaning, and four engines that kept reading it
Symptom. The "final live config" backtest reported a median terminal wealth of 0x — every sampled path lost everything, half of them to a bust. That looked like a verdict on the strategy. It was a verdict on a column.
Root cause. On 2026-09-05 a migration moved ret_24h up 24 rows:
rn[24:] = r[:-24] # docstring: "→ historical returns"
After that write, the column is trailing 24h momentum — something you knew
before entering. Every engine that consumed it as the forward 24h return kept
subtracting fees and calling the result net PnL. Four scripts do this
(bt_final_live.py, bt_2pos_circuit.py, bt_withdraw_design.py,
bt_path_enum.py); the first one was the "final configuration".
Evidence that the column is now backward-looking. On the current
proxy_1h_v5.csv, of 2,011 sampled rows, 2,008 match c[i]/c[i-24]-1
(the past) and 3 match the future. The author's own 2026-09-15 note already said
ret_24h correlates +1.0000 with the past 24h — the migration was known; the
consumers were never revisited.
Same script, two semantics (differs by one line, applied at load time):
| Data in the column | Median terminal wealth | Loss paths | Busts | Ended >=1000x |
|---|---|---|---|---|
| Current CSV (trailing return used as PnL) | 0x | 100% | 50.0% | 0% |
| Rebuilt as forward 24h return | 1,760x | 16.7% | 16.7% | 65% |
Neither number is evidence about the strategy. The first is arithmetic on the wrong sign of time; the second is arithmetic on the right sign of time with a backtest that still assumes you can fill at the close.
The second correction: 1,760x is not an edge either. Three experiments on the rebuilt column:
- Event quality. 487 events (DIV 217 / BRK 180 / WICK 90). Forward 24h mean +3.77%, median +2.45% after 29 bps. But median MAE is −4.55% and P(MAE <= -16%) = 8.8% — the path, not the endpoint, decides survival.
- Liquidation. At 5x, marking only at the close gives a median of 39.8x with no busts. Add intrabar liquidation and the same trades give a median of 0.02x, with 17% busted. At 3x: 0.57x. At 1x: 601x. The 5x result was a fiction produced by ignoring the wick.
- Concurrency. Half the events overlap another coin's event; median 2 open at once, p90 4, max 10. The old script compounded serially at full size. At 1x with a position cap over three years: 1.71x (K=1) · 5.28x (K=3) · 3.94x (K=5) · 2.26x (K=10).
One more caution: the "216 paths" in the original write-up are the 216 permutations of six year-blocks. The independent sample size is six, not 216; percentages computed across them are pseudo-replication.
Bottom line. A column header is not a definition. When a data migration changes the direction of time in a column, every consumer must be re-run, not just the one that looks important. And fixing the direction gets you back to arithmetic — it does not get you to an edge: leverage without intrabar liquidation, and compounding without a concurrency cap, will manufacture one.
Caveats (scope). Audited on one machine's archive; no venue replay, no order
book, no funding. The pre-migration file
(proxy_1h_v5.csv.bak_0905_poison) is no longer on disk, so "the column used
to be forward-looking" is inferred from the migration script, the engine
semantics, and three historical magnitudes — not read directly from the old file.
Fees used: 29 bps per event round trip; slippage beyond that is not modelled.