Does the best two-measure positioning rule beat holding BTC on blind data?
FRAME
- Subject: BTCUSDT USDT-M perpetual, hourly, Binance Vision (1h klines, fundingRate, metrics).
- Why it matters: LB6's edge seems to come from positioning data beyond price, found by a threshold sweep; this is the same search shape on another positioning source, with the search locked and its luck measured.
- Prior work: none on this data.
- Scoreboard: Calmar (total % / max drawdown, hourly mark-to-market).
- Scope: locked sweep, N = 3,888; dev 2020-09-02 to 2024-12-31; holdout 2025-01-01 to 2026-10-31, used once if dev gates pass. Locked in
docs/bets/2026-11-18-positioning-sweep.md, commit77ef138(top 5 dev candidates seen in a timing run before locking). - Terms: features funding_rate, oi_chg_24h, global_ls, taker_buy_share; rule = f1 op q1 and f2 op q2, one side, held 24 h after the last true hour, fee 0.05%/side, funding included; thresholds = dev percentiles 10..90; placebo = same sweep with features shifted 30 days or more against prices.
BET
The best positioning rule found on dev is more efficient than holding BTC on blind data.
- Adjective: more efficient
- why: a rule that trades part of the time must earn more per unit of risk than holding, or it adds nothing.
- Measured by: Calmar on the holdout; higher wins.
- why: same measure as all reports here.
- Because: positioning extremes are widely reported to precede reversals.
- Falsified if: dev winner <= placebo p95; or neighbours' median Calmar <= 0; or holdout Calmar <= max(buy-and-hold, 0).
KILL
Winner on dev: long when oi_chg_24h >= -2.79 (p20) and global_ls <= 1.716 (p60); 163 trades, total +396.0%, max drawdown 31.0, Calmar 12.79. Holding BTC on dev: total +292.1%, Calmar 2.32.
| test | falsified if | result | status |
|---|---|---|---|
| dev: winner Calmar vs placebo p95 | <= 9.27 | 12.79 | pass |
| dev: median Calmar of threshold neighbours | <= 0 | 9.11 | pass |
| holdout: winner Calmar | <= 0.00 (max of buy-and-hold, 0) | -0.38 | FAIL |
Placebo best Calmar over 100 runs: median 5.77, 95th percentile 9.27, max 10.11; real winner above 100 of 100.
Holdout 2025-01-01 to 2026-10-31: rule total -19.9%, max drawdown 52.3, Calmar -0.38, 60 trades; holding BTC total -1.5%, max drawdown 69.3, Calmar -0.02.
Post-hoc diagnosis (after the result): share of hours in a position, and BTC's summed hourly return while in vs out of a position: dev: in market 68%, BTC return in +443.5%, out -151.4%; holdout: in market 77%, BTC return in -8.0%, out +6.4%.
Threshold neighbours of the winner:
| q1 | q2 | trades | total_pct | calmar |
|---|---|---|---|---|
| p10 | p50 | 129 | +368.3 | 8.92 |
| p10 | p60 | 132 | +409.6 | 9.47 |
| p10 | p70 | 130 | +416.6 | 7.33 |
| p20 | p50 | 156 | +356.7 | 9.65 |
| p20 | p70 | 164 | +395.3 | 8.50 |
| p30 | p50 | 189 | +344.3 | 9.30 |
| p30 | p60 | 205 | +385.7 | 12.43 |
| p30 | p70 | 213 | +365.6 | 8.23 |
Top 10 candidates on dev (>= 71 trades); all 3,888 in candidates.csv:
| side | f1 | op1 | q1 | f2 | op2 | q2 | trades | total_pct | max_dd_pts | calmar |
|---|---|---|---|---|---|---|---|---|---|---|
| long | oi_chg_24h | >= | p20 | global_ls | <= | p60 | 163 | +396.0 | 31.0 | 12.79 |
| long | oi_chg_24h | >= | p30 | global_ls | <= | p60 | 205 | +385.7 | 31.0 | 12.43 |
| long | oi_chg_24h | >= | p40 | global_ls | <= | p60 | 245 | +376.9 | 30.9 | 12.19 |
| long | funding_rate | <= | p30 | global_ls | <= | p30 | 96 | +200.4 | 17.2 | 11.65 |
| long | global_ls | <= | p40 | taker_buy_share | >= | p70 | 135 | +339.8 | 30.0 | 11.33 |
| long | global_ls | <= | p40 | taker_buy_share | >= | p80 | 142 | +321.8 | 28.6 | 11.24 |
| long | oi_chg_24h | <= | p60 | global_ls | <= | p40 | 201 | +335.4 | 30.0 | 11.17 |
| long | funding_rate | <= | p20 | global_ls | <= | p30 | 95 | +176.5 | 16.2 | 10.88 |
| long | global_ls | <= | p70 | taker_buy_share | <= | p20 | 177 | +393.2 | 36.6 | 10.75 |
| long | oi_chg_24h | >= | p60 | global_ls | <= | p60 | 320 | +339.3 | 31.7 | 10.69 |
UPDATE
- Status: FALSIFIED.
- Next bets:
positioning-sweep-stable(Does a positioning sweep that requires gains in most dev years, not one episode, find a rule that lasts?)