Strategies

Contrarian BTC long on extreme negative funding spikes

Outcome: Abandoned

BtcNegativeFundingExtremeContrarianLong

Outcome Summary

This was a well-researched, honestly-caveated hypothesis: an empirical study of 7,166 funding settlements showed BTC systematically squeezes up after extreme-negative funding, so the strategy bought BTC on sub-percentile funding prints and held ~72h, using a rolling percentile to stay adaptive as funding compressed across eras and collecting the negative funding while waiting. The backtest confirmed the effect existed but also confirmed the hypothesis's own regime-risk caveat: the +116% return came entirely from 2020 and 2023, while 2025 and 2026 were sharply negative and rolling Sharpe fell from +8 to -7, leaving a Sharpe of 0.378 with a CI straddling zero and much of the gain attributable to plain BTC beta. The analyst abandoned it at backtest review as a regime-concentrated, decayed edge — funding has compressed an order of magnitude and the ~72 extreme events cluster in early capitulations — judging that with the holdout in the negative 2026 regime, optimization would only overfit noise and fail the robustness gates.

Strategy report

Hypothesis

Empirical study of 7,166 Binance BTCUSDT 8h funding settlements (2020-2026) shows BTC systematically PUMPS after negative funding spikes, not dumps (short-squeeze / mean-reversion). Bottom-1% funding events (fr < ~-0.017%) are followed by +1.30% mean at 8h, +2.25% at 24h, +4.10% at 72h with a 72% 72h win rate; Spearman corr(funding, fwd return) is negative across all horizons. Signal: go LONG BTC when funding prints below its ~1st percentile at settlement, hold 24-72h, exit on time-stop or mean-reversion target. Route to a low-fee futures/perp venue (HYPERLIQUID ~0.09% or Binance USD-M ~0.10% round-trip) since it needs shorts-free leverage and the edge is funding-based. Caveats to validate: only 72 extreme events in 6.5y likely cluster around 2020/2022 capitulation (regime risk); median move ~60% of mean so size to the median; needs walk-forward + holdout to confirm out-of-sample. Entry uses funding known AT settlement (no lookahead).

I validated the hypothesis against the real Binance BTCUSDT funding series (7,029 settlements, 2020-01 to 2026-05) and 8-HOUR catalog bars before coding, and the headline result required one substantive change. THE FIXED THRESHOLD IS DEAD: funding below -0.017% fired 40 times in 2020, 20 in 2021, 11 in 2022, and ZERO in 2023, 2024, 2025 and 2026. The funding regime compressed by more than an order of magnitude (2020 min -0.300%, 2025 min -0.0122%), so all 71 of the hypothesis's extreme events are a 2020-2022 capitulation artifact. This confirms the hypothesis's own flagged caveat ('likely cluster around 2020/2022 capitulation') completely, and it means a fixed cutoff would have produced ZERO trades in the 365-day Layer-3 sandbox and been abandoned immediately. The fix is the one the hypothesis itself points at ('below its ~1st percentile') plus L5: make the percentile ROLLING over the instrument's own recent history, which is scale-invariant and keeps the mechanism firing in every era (~25 events/year). Two honest deviations, both measured, not tuned. (1) PERCENTILE 5 NOT 1: with a rolling threshold the 1st percentile gives +0.72% net per trade at t=1.34, while the 5th gives +0.99% at t=2.19 with a broad plateau (P=5-10 x window=180-270 all positive at t~2.0-2.2), so 1% is the thin, noisy end rather than the sweet spot. The hypothesis's +4.10%/72%-win figure does NOT reproduce under a regime-adaptive threshold; it is specific to the fixed cutoff that only selects 2020-2022. (2) NO TAKE-PROFIT OR TIGHT STOP: adding a 4% target / 3% stop collapsed the edge from +1.31% to +0.08% per trade because the squeeze is slow and choppy and gets stopped out on the way up, so the exit is a time stop plus a wide 12% disaster brake. Robustness of the shipped defaults over the full 2020-2026 history: n=154 non-overlapping trades, +0.994% average net per trade after the 0.10% USD-M round trip (well clear of the 0.15% floor), 57% win rate, PF 1.65, t=2.19, positive in 6 of 7 calendar years (only 2025 negative at -0.44%), and still +0.815% excluding the outlier 2020 vintage. Hold sensitivity is also a plateau (24h +0.50%, 72h +0.99%, 96h +1.09%). I picked mid-plateau values rather than the grid maximum. Implementation notes: 8-HOUR bars align exactly with the funding cadence so each settlement is one bar; funding is read at-or-before the bar timestamp with a 12h staleness tolerance (>= the 8h interval per the data spec) so there is no lookahead; and the funding dict is parsed from BOTH its int and str key forms because the Layer-3 sandbox JSON-serialises the supplementary blob into its subprocess, which stringifies every key -- I verified a round trip recovers 400/400 settlements where an int-only reader recovers 0. Venue is Binance USD-M futures rather than spot even though the strategy is long-only: the edge is funding-based, and a long in negative funding actively COLLECTS the funding payment while it waits, which the engine accrues on this venue. Leverage 1.5 is genuinely consumed in position_size (75% of equity notional, not 50%), and combined with the 12% stop bounds the worst single-trade loss near 9% of equity.

Regime-concentrated, decayed edge that will not clear the post-optimization robustness gates. Sharpe 0.378 with a 95% CI [-0.205, 1.042] straddling zero, profit_factor 1.14 (below the 1.2 meaningful line), and information_ratio -0.58. The 116% total return is entirely from 2020 (+68%) and 2023 (+46%); the recent regime is deeply negative (2025 -23.3%, 2026 -7.1%) with rolling Sharpe collapsing from +8 (2021) to -7 (2026) — the premise has decayed exactly as the hypothesis itself warned ('only 72 extreme events likely cluster around 2020/2022 capitulation; regime risk'), consistent with Binance BTC funding compressing by an order of magnitude and shrinking the short-squeeze edge. Being long-only BTC, much of the headline is market beta (beta 0.19, correlation 0.49, alpha only ~3.3%), and max_drawdown is 36% (CI high 68%) with return_kurtosis 12.8 (capitulation-outlier-driven). avg_trade_return_pct 0.72% clears the fee floor, so this is regime-decay/overfit, not fee-fragility. Entering the optimizer at Sharpe 0.38, best-of-225 selection would inflate it on noise and the 15-day holdout falls in the sharply-negative 2026 regime — a near-certain deflated-Sharpe/holdout failure, with no robust parameter region to tune toward when the mechanism has stopped working in the modern regime. Abandon at BACKTEST_REVIEW rather than spend 2 hours optimizing.

Implementation

Long-only contrarian strategy on BTCUSDT.BINANCE USD-M perpetual (8-HOUR bars matching the funding cadence, 1.5x leverage). Each bar it reads the funding rate settled at or before the bar's own timestamp and returns the negated z-score of that rate against its own rolling window as the continuous signal (positive = funding unusually negative = shorts crowded and paying = bullish). The entry gate is the hypothesis's rule expressed as a ROLLING percentile: go LONG when funding prints at or below the funding_pctile-th percentile of the last pctile_window settlements. Exits are (1) a wide 12% catastrophic stop tested intrabar against the bar's low, and (2) a hold_bars time stop (9 x 8h = 72h) derived from calendar timestamps. There is no take-profit and no tight stop, by measurement rather than by omission. If funding is missing or stale the strategy stays flat and never substitutes a price-only proxy. Sizing is a fixed 50% x leverage (= 75%) of equity notional per trade, one position at a time with no averaging down.

Verification Results

Sandbox net-negative (−15.9%, −0.488%/trade, PF 0.61) but this overlaps the developer's own flagged 2025 bad year (−0.44%/trade), consistent with their model — not a code defect or engine-vs-study divergence. Full-history edge (+0.994%/trade, t=2.19, 6/7 positive years) is unverified OOS.

Regime-concentration risk: the hypothesis's +4.10% figure is a 2020-2022 capitulation artifact; the rolling-percentile fix keeps it firing but the +4.10% does not reproduce (+0.994% is honest). Recent-years walk-forward/holdout is decisive.

Tail/leverage risk for Risk Officer: 0.75x notional at 1.5x leverage, skew −2.29, worst sandbox trade ~11.5% of equity (gap through the 12% stop). Hypothesis advised median-scaled sizing; shipped fixed 0.75x does not.

Backtest Review

Sharpe: 0.38
Total return: 116.04%
Max drawdown: 36.02%
Trades: 214
Win rate: 52.3%
Profit factor: 1.14

Well-researched, honestly-caveated hypothesis; clean no-lookahead funding construction; avg_trade_return_pct 0.72% clears the fee floor

Reasonable trade count (214) and capacity ($9.2M); rolling-percentile gate is regime-adaptive by design

Sharpe 0.378 with 95% CI [-0.205, 1.042] straddling zero; profit_factor 1.14 below the 1.2 line; information_ratio -0.58

Edge entirely concentrated in 2020 (+68%) and 2023 (+46%); recent regime deeply negative (2025 -23.3%, 2026 -7.1%) with rolling Sharpe collapsing from +8 to -7 — the premise has decayed exactly as the hypothesis warned

Long-only BTC: beta 0.19, correlation 0.49 — much of the 116% headline is market beta, alpha only ~3.3%

max_drawdown 36.0% (CI high 68%); return_kurtosis 12.8 — outlier/capitulation-driven

Funding has compressed by an order of magnitude, shrinking the short-squeeze edge; only ~72 truly extreme events cluster in early capitulations (small effective sample)

15-day holdout falls in the negative 2026 regime — near-certain post-optimization failure

Outcome Summary

A well-documented empirical edge can be genuinely real yet untradeable going forward if it has decayed — the negative-funding short-squeeze depended on extreme events that clustered in 2020/2022 capitulations, and as funding compressed an order of magnitude the edge collapsed (rolling Sharpe +8 to -7), so a rolling-percentile gate keeps the mechanism firing but cannot restore an effect the modern regime no longer produces.

It was abandoned at the BACKTEST_REVIEW gate: a Sharpe straddling zero, profit factor below 1.2, and an edge that has decayed exactly as the hypothesis itself warned — Binance BTC funding has compressed by an order of magnitude, shrinking the short-squeeze edge, with only ~72 truly extreme events clustered in early capitulations. With the 15-day holdout falling in the sharply-negative 2026 regime, optimization would inflate a Sharpe-0.38 signal on noise and near-certainly fail the deflated-Sharpe/holdout gates, so it never reached optimization.

A long-only contrarian BTC strategy on BINANCE USD-M (8h bars) that buys BTCUSDT when the perpetual funding rate settles below its rolling ~1st-5th percentile (crowd crowded short and paying longs, a short-squeeze setup that empirically pumps rather than dumps), holding ~72 hours to a time stop with a wide catastrophic stop, and collecting the negative funding while it waits — grounded in an empirical study of 7,166 funding settlements showing BTC pumps +4.1% mean at 72h after bottom-1% funding events.

The edge was real historically but regime-concentrated and decayed: total return +116%, avg_trade_return_pct 0.72% (above the fee floor), 214 trades, but Sharpe only 0.378 (CI [-0.205, 1.042] straddling zero), profit factor 1.14, information ratio -0.58, max drawdown 36% (CI to 68%). The return came entirely from 2020 (+68%) and 2023 (+46%), while the recent regime was deeply negative (2025 -23.3%, 2026 -7.1%) with rolling Sharpe collapsing from +8 to -7, and as a long-only BTC book much of the headline was market beta (beta 0.19, alpha only ~3.3%).

Backtest and paper results are hypothetical. Trading involves risk of loss.