E-25. Measuring detection power of candle proxies vs real liquidation clusters (2026-07-16)
Nature: instrumentation, not strategy validation. It fills the blank left in E-10 as "proxy validity untested". No preregistered gate — a measurement, not a verdict.
- Data: COIN-M liquidations over 472 days, 53,398 records (BTCUSD_PERP) × USDT-M BTCUSDT 1m, 687,120 bars. Comparison window 2023-06-25 ~ 2024-10-14 (entirely inside train — no boundary breach).
- Proxy (cascade_proxy default parameters: volume 8x + 1-minute -0.5% + recovery wick 0.2%) fired: 137 times out of 687,120 bars (0.02%).
- Results:
| Cluster definition | Clusters | recall | precision | Null base rate | lift |
|---|---|---|---|---|---|
| ≥3 events / gap≤1 min | 3,460 | 4.7% | 93.4% | 4.4% | 21.1x |
| ≥3 events / gap≤2 min | 3,422 | 3.5% | 87.6% | 5.2% | 16.8x |
| ≥5 events / gap≤1 min | 1,648 | 8.4% | 90.5% | 2.3% | 39.1x |
| ≥5 events / gap≤2 min | 1,787 | 6.2% | 86.1% | 3.0% | 29.0x |
| ≥10 events / gap≤1 min | 492 | 18.9% | 69.3% | 0.8% | 89.5x |
| ≥10 events / gap≤2 min | 630 | 13.2% | 71.5% | 1.2% | 60.1x |
- Reading: the lift, precision divided by the null base rate (the share of the whole period covered by cluster windows), is 21~90x. That means 20~90x more accurate than picking random times, confirming it is not a base-rate trap (an illusion from clusters covering a lot of time).
- Conclusion 1 — the proxy was not inaccurate; it was insensitive. When it fired, a real cascade almost always existed (precision 69~93%). Instead it missed 81~95% of real cascades. So the cause of the E-10 rejection narrows from "it traded noise" to "the rebound after confirmed real cascades did not exceed costs" — consistent with the E-10 post-mortem ("by the time the recovery wick shows, the rebound is already spent").
- Conclusion 2 — the unexplored space is large. Of 3,460 clusters, the proxy saw 137. Most cascades, which leave no fingerprint in candles, are reachable only with real liquidation data, and that is the main stage for the branch hypothesis (planned D-036).
- Limits [explicit]: (a) a cross-market comparison of COIN-M liquidations ↔ USDT-M candles — it rests on the assumption that cascades propagate to both markets simultaneously (this measurement is itself indirect evidence for that assumption). (b) It is a throttled sample, so cluster size is a "proxy for duration", not the real liquidation size. (c) Recall is both a flaw of the proxy and the "size of the opportunity" — the tradable share of what was missed is unknown.
Source and migration evidence
Source: backtest/docs/experiment-log.md#E-25
Original SHA-256: 069c4aaaf53a20930ca8d94ba6365b4ce2b6bc560171d662a4c179ea55a6857c
Raw data and account information are not part of the public package. The original results do not replace statistical or scientific revalidation. The English text is a translation; numbers, tables and charts are carried over unchanged.