Is a BTC grid whose side follows LB6 more efficient than a plain grid on the holdout?
FRAME
- Subject: BTC grids modelled on Hummingbot's GridExecutor (
lib/grid.py) on Binance Vision BTCUSDT USDT-M perpetual 1-minute bars; LB6 Config B simulated tradelist for direction. - Why it matters: pak Yon's request: a grid strike whose side is set by LB6, judged by volume, with profit from LB6's direction.
- Prior work: lb6-volume-baseline (MEASURED): LB6 net direction 58% of the time before 2026-02-01, ~30 direction changes per 30 days. grid-neutral (MEASURED): before 2026-02-01 neutral grid -2.1%, Calmar -0.20; its BUY half (long-only) Calmar -0.48.
- Current best: the better plain grid on the same window by Calmar: neutral (0.65).
- Scoreboard: pak Yon judges by volume (rebates) with profit from LB6's direction; how they combine is [not confirmed] (w-zb8).
- Question: does following LB6's side turn grid volume into a better return per unit of risk than a plain grid?
- Scope: HOLDOUT 2026-02-01 to 2026-08-26 (206 days), its one decisive use. BTC moved -0.2%. LB6 was net long 33%, net short 16%, flat 51% of the time. LB6's own thresholds may have been set on data including this window; the grid choices were not. Parameters: step 0.5% (median BTC 1h range before 2026-02-01: 0.46%); half-range 3%, 13 levels (median 24h range 2.63%; LB6 median hold 24h); take-profit one step; stop one step beyond the far end; at most 2 resting open orders; maker fee 0.02%, taker fee 0.05%; 1,000 USDT capital, leverage 1; re-anchors at the current price when an executor ends.
- Terms: LB6-directed = BUY grid while LB6 net long, SELL grid while net short, none while flat; a direction change closes the grid at market and restarts. Break-even maker fee = the maker fee at which total would be zero.
BET
A grid whose side follows LB6 is more efficient than the better plain grid.
- Adjective: more efficient
- why: the idea is that LB6's direction turns a grid's volume into profit instead of loss; efficiency weighs the profit against the risk taken.
- Measured by: Calmar = total net % / max drawdown (1-minute mark-to-market); higher wins.
- why: same measure as the LB6 reports and on pak Yon's scoreboard [not confirmed for grids]; it counts open grid inventory.
- Because: before 2026-02-01 the grid's side decided its sign (SELL half +2.6%, BUY half -4.8%, both in % of the 1,000 USDT total).
- Falsified if: Calmar(LB6-directed) <= max(Calmar(neutral), Calmar(long-only)). Locked in
docs/bets/2026-11-18-grid-strike-lb6.md, commit4bb3c96(2026-11-17 17:23 UTC = 2026-11-18 00:23 WIB; bet files are dated in WIB), written before any grid ran on the holdout.
KILL
All three grids ran once on the holdout with identical parameters.
| test | falsified if | result | status |
|---|---|---|---|
| Calmar, LB6-directed | <= 0.65 (neutral) | 2.22 | pass |
Also observed, not deciding:
| variant | total_pct | max_dd_pts | max_dd_at | calmar | volume_usd | round_trips | executors | ends_tp_stop_dir | breakeven_maker_fee_pct |
|---|---|---|---|---|---|---|---|---|---|
| LB6-directed | +8.88 | 4.00 | 2026-03-29 | 2.22 | 136,575 | 518 | 152 | 33/13/106 | +0.1021 |
| neutral | +4.87 | 7.53 | 2026-08-25 | 0.65 | 231,801 | 1802 | 205 | 109/94/0 | +0.0463 |
| long-only | +4.10 | 10.48 | 2026-06-24 | 0.39 | 231,590 | 886 | 107 | 57/49/0 | +0.0422 |
executors counts every grid started; ends counts those that ended before the window closed (a grid still running at the end is in the first number only). Max drawdown dates are the trough; LB6-directed fell from +4.48 (2026-02-28) to +0.48 (2026-03-29).
Post-hoc check (run after the result was seen; not part of the bet): same on/off timing as LB6 with the side reversed: total -15.00%, Calmar -0.86. Same timing with a random side per LB6 stretch, 10 seeds: Calmar median -0.25 (min -0.84, max 0.89); total median -1.58%; 0 of 10 reach LB6's Calmar; 1 of 10 beat the neutral grid.
Post-hoc, sensitivity to the end of the window: up to 2026-08-18 (before BTC's last-week rally), LB6-directed total +9.04%, max drawdown 4.00, Calmar 2.26; neutral total +11.47%, max drawdown 5.55, Calmar 2.07. Most of the margin on the full window comes from the final week, where the neutral grid's SELL half lost.
An independent re-implementation from the same written spec and the Hummingbot source (subagent, no shared code, ~/.cache/claude-work/xcheck-grid/grid.py) reproduced all three variants' total, max drawdown, Calmar and volume to the shown precision.
Volume per 30 days per 1,000 USDT: LB6-directed 19,890; neutral 33,758; long-only 33,727.
UPDATE
- Status: SURVIVED.
- What it means: LB6-directed total +8.88% vs neutral +4.87%; max drawdown 4.00 vs 7.53; volume 136,575 vs 231,801 USDT.
- Caution: the margin over the neutral grid is fragile; up to 2026-08-18 the two were close (2.26 vs 2.07) and neutral had the higher total. LB6-directed traded 59% of neutral's volume.
- Open: pak Yon's scoreboard for this grid (volume vs profit); grid parameters were set once and not explored; a parity check of
lib/grid.pyagainst a Hummingbot backtest (w-2vl).