Skip to content
All strategies

Strategies

LINK Leveraged-Demand Continuation: Long for a Fixed 5 Days When the 72h Mean Funding Rate Is in the Top 20% of Its Own Trailing 90 Days (LINKUSDT.BINANCE USD-M, 4H Bars, Long-Only With a Flat State, ~29% Time in Market)

Outcome: Abandoned

LinkFundingDemandContinuation

Outcome Summary

LinkFundingDemandContinuation bet that top-quintile LINK funding signals continued leveraged demand and held a fixed 5-day long, flat about 71% of the time. With its pre-registered parameters it returned +106% over 136 trades with a 1.90% average trade, but the Sharpe of 0.60 was not significant, drawdown was 44%, and PF and win rate fell short of the pre-test. The optimizer drifted to a much looser entry threshold with higher sizing and was liquidated at -100%. The holdout Sharpe of -0.69 on 64 trades was a hard failure, so the analyst abandoned it after one iteration as overfit. The analyst also flagged a suspected MTM artifact in daily_returns (paired -X/+X swings while flat) for engine review; that concern has not been verified.

Strategy report

Hypothesis

Single-leg, single-instrument, long-only trade on the LINKUSDT.BINANCE USD-M perpetual. The trigger is a funding-rate event, and every exit is a fixed clock. At each 4H bar close, compute F = the mean of Binance funding prints over the trailing 72 hours, read point-in-time with supp_window('funding_rate') and never a nearest-timestamp match. Compute Q = the 80th percentile of F over the trailing 540 4H bars (90 days), also point-in-time. If flat and F > Q, buy at the next bar and hold exactly 30 bars (5 days),……Show moreShow less

Single-leg, single-instrument, long-only trade on the LINKUSDT.BINANCE USD-M perpetual. The trigger is a funding-rate event, and every exit is a fixed clock. At each 4H bar close, compute F = the mean of Binance funding prints over the trailing 72 hours, read point-in-time with supp_window('funding_rate') and never a nearest-timestamp match. Compute Q = the 80th percentile of F over the trailing 540 4H bars (90 days), also point-in-time. If flat and F > Q, buy at the next bar and hold exactly 30 bars (5 days), then exit. The strategy is FLAT about 71% of the time. Why the sign is continuation, not a contrarian fade: across 10 liquid Binance perps (2020-2026, 4H), 5-day forward returns after F entered its own top decile exceeded the unconditional 5-day drift on 8 of 10 names. The crowded-short squeeze long and the funding-reset long showed no excess over drift and were discarded. Pre-test on LINK (non-overlapping trades, 0.10% round trip deducted, 1x notional): 141 trades, mean +1.85%/trade, median +0.93%, win rate 56%, PF 1.59, 29% time in market, calendar Sharpe about 0.74; the last 20% of trades (29 trades) averaged +1.09%. Selection choices, stated explicitly: LINK was chosen from the 10-name pass (count about 10 trials); Binance USD-M is required for the funding feed and the 6-year history; only 80/90 percentile and 30/42-bar holds were tried. Caveat: the signal behaves like amplified beta, and 2022 was negative on every name. Risk: 0.5x equity notional, a single position, 15% disaster stop, no take-profit and no trailing stop.

This follows the hypothesis: a funding-level trigger (continuation, not a fade), a fixed 30-bar clock exit, one position at 0.5x notional and a 15% disaster stop. There is no trailing stop and no real take-profit. The base bracket requires take_profit_pct > 0, so the take-profit leg is set to +500%, where it can never trigger over a 5-day hold. The stop fires intra-bar at the venue instead of only at bar closes. Funding is read through the base's point-in-time sorted arrays: only prints at or before the bar close……Show moreShow less

This follows the hypothesis: a funding-level trigger (continuation, not a fade), a fixed 30-bar clock exit, one position at 0.5x notional and a 15% disaster stop. There is no trailing stop and no real take-profit. The base bracket requires take_profit_pct > 0, so the take-profit leg is set to +500%, where it can never trigger over a 5-day hold. The stop fires intra-bar at the venue instead of only at bar closes. Funding is read through the base's point-in-time sorted arrays: only prints at or before the bar close are used, with no nearest-timestamp matching. The 90-day F history is rebuilt from bar-grid timestamps on the first bar or after any gap, so it is identical across restarts and needs no 540-bar warmup, since the base caps _bars at 500. Per-bar cost is O(log n) plus a 540-element percentile. If the funding feed is missing, the strategy logs data_unavailable and stays flat rather than falling back to a price-only proxy. At a 29% time-in-market and a 30-bar hold, expect about 40-50 trades over a 20k-bar sandbox window. Caveat, as the hypothesis states: the signal acts like amplified beta.

Failure code: overfit (holdout failure), classified from the first measured report: avg_trade 1.90% vs 0.15% floor (not fee_edge), PF 1.24, Sharpe 0.60 (CI -0.22 to 1.32). That report missed its own DD and PF predictions. Holdout Sharpe -0.69 on 64 trades (hard gate). The optimized config drops the top-quintile premise and was liquidated at -100%. The edge decays as the funding threshold becomes more extreme, so the premise is not supported on LINK. The hypothesis's own 'amplified beta' caveat was borne out. Do not re-propose single-name LINK funding-percentile continuation. Only a 10-name panel at fixed pre-registered thresholds, with sizing not optimized, would be a fair restatement. Engine review (unverified): paired -X/+X MTM swings while flat and a +3558% day point to an MTM folding artifact. notional_frac should not be an optimizer free parameter.

Implementation

Long-only LINKUSDT.BINANCE USD-M perp on 4H bars. On each bar close it computes F, the mean of the Binance funding prints stamped in the trailing 72h, point-in-time. It also computes Q, the 80th percentile of F over the trailing 540 4H bars (90 days). If flat and F > Q, it buys 0.5x equity notional at 1x leverage and exits after exactly 30 bars (5 days), timed from the position's ts_opened. A venue-managed 15% disaster stop is placed as a bracket stop-market. The bracket's take-profit leg is set to +500%, which disables it in practice. The signal is (F - Q) in bps.

Verification Results

No change is required for fidelity. If the drawdown floor fails at the full backtest, lower notional_frac rather than changing the signal.

Sizing is 0.5 x current equity x leverage, recomputed from equity at every entry. The worst case is 0.5 x 1.0 leverage x 1 leg = 0.5x equity gross notional, well under the venue cap. With the 15% bracket stop, planned risk per trade is about 7.5% of equity, more than the 1-2% guideline. That matches the hypothesis's stated risk profile (0.5x notional, 15% disaster stop), so it is not a defect. Gap-through fills on the stop-market can exceed the 7.5%.

None needed. Optionally, log the number of backfills to spot data gaps.

Any bar-grid discontinuity (ts - last_ts != interval) triggers a full backfill of 539 historical F values. Each value costs a bisect plus a 9-print mean. The cost is bounded and point-in-time, but a gappy catalog would repeat it on every gap. That is fine for contiguous 4H data.

Keep it, and keep it out of the optimizer's search space.

take_profit_pct is set to 5.0 (+500%) only to satisfy the base's bracket requirement. It effectively disables the take-profit, as the hypothesis intends (no TP). It is documented in the config rationale.

The sandbox shows +82% total return over 140 trades, PF 1.17, win rate 45.7% and avg trade +1.77% of notional. The pre-test claimed PF 1.59 and 56% wins. Max DD is 44% at 0.5x notional, and the bootstrap max-DD CI runs 68-99%. The 'calm' vol tercile lost 79.6%, beta to buy-and-hold is 0.18 with correlation 0.53, and information ratio is -0.65. This agrees with the hypothesis's own caveat that the signal is amplified beta. At the full backtest, check first: (1) whether excess return over the unconditional 5-day drift survives; (2) 2022 and the calm regime, where most of the drawdown sits; (3) whether the drawdown clears the hypothesis's pre-registered floor at 0.5x notional. Ledger reconciliation is fine: 1.77% x 140 x 0.5 is about +124% simple, against +82% compounded, which is within tolerance. The avg trade is well above the 0.15% fee floor.

Backtest Review

Sharpe
0.60
Total return
106.04%
Max drawdown
44.38%
Trades
136
Win rate
44.9%
Profit factor
1.24

The trades implement the hypothesis: 136 long-only trades, nearly all 5d 0h fixed-clock exits, 28.75% exposure (29% pre-registered), and a working 15% disaster stop (about -15% exits in Nov 2022 and Mar 2024). There are 0 shorts.

Per-trade economics are far above the USD-M fee floor. Report avg_trade_return_pct is 1.90%; the mean trade pnl_pct recomputed from the detail file is 1.32%, net of funding paid. Commission is only 2.08% of gross and impact 8.2%.

Adequate sample over 6.2 years: 136 trades, 1865 funding events loaded, no data gaps. Five of seven calendar years are positive by trade mean (2021 +5.3%, 2022 +0.9%, 2023 +1.4%, 2024 +2.1%, 2025 +1.3%). The hypothesis's warning that 2022 would be negative did not come true on LINK.

End unrealized is 0%, so the +106% headline is fully realized.

The pre-registered profile did not reproduce. The pre-test claimed PF 1.59, win rate 56% and median +0.93%/trade. The backtest shows PF 1.24, win rate 44.9% and median -0.34%. QA's concern is partly borne out: the edge is thinner and more right-skewed than the pre-test said.

Max drawdown is 44.4%, recomputed from the trade-level equity curve, at 0.5x notional. That is far above the 20% futures default. The drawdown takes 389 days to recover, and the bootstrap DD CI runs 53-96%.

The base Sharpe of 0.60 (CI -0.22 to 1.32, PSR 0.81) is not significant. Beta is 0.17 and alpha +4.6%/yr, which is small: this is mostly amplified-beta timing, as the hypothesis admits.

2020 (-3.2%/trade across 22 trades) and 2026 YTD (-1.9% across 8) are negative.

Suspected metrics artifact, mechanism not independently verified. daily_returns contains repeated paired phantom dips that net to zero, e.g. -17.83%/+21.70% on 2022-01-22/23, 2022-03-13/14, 2022-05-02/03 and others, and -28.64%/+40.13% four times in 2024. They fall on dates with no open trade: the equity_curve is flat (130358.11) through 2022-01-22..27. These inflate volatility and CVaR (13.8%) and distort the regime split (calm = -61%) and the rolling Sharpe. The final 2026-04-08 -22.09% value is also suspect. The calendar Sharpe and regime/tail metrics should be read with caution. The trade ledger and equity curve look clean. Recommend engine review of the daily_returns construction.

Analysis

First measured report (pre-registered params): 136 trades, avg_trade_return_pct 1.90% vs 0.15% USD-M floor (12.7x), so not fee_edge.

Not buy-and-hold beta by L170: beta 0.17, DD 44% at 29% exposure vs LINK buy-and-hold DD of about 90%.

Smooth sensitivity at the pre-registered point; CPCV PBO 0.06.

HARD FAIL: holdout Sharpe -0.69 on 64 trades, 2.1 s.e. below WF-OOS (consistent_with_oos=false). Base run also -29% in 2026 YTD.

Optimizer left the hypothesis: entry_percentile 42 (long whenever funding is above its median), notional 0.745. Result: liquidated 2024-10-09, total_return -100%, PF 0.92, avg_position 369%, metrics_reliable=false.

Pre-registered config missed its own predictions (DD 44% vs 25%, PF 1.24 vs 1.35, WR 0.45 vs 0.50). Sharpe CI straddles 0.

DSR 0.63; fails programme FDR. The edge weakens as the funding threshold becomes more extreme (96th pct Sharpe about 0.2), which contradicts the premise.

Unverified engine concern: recurring identical paired -X/+X MTM days while flat, and a +3558% day on 2022-05-03.

Benjamini-Hochberg at q=0.10 over 322 programme candidates keeps 6. A candidate that does not survive here is not distinguishable from the programme's own noise, however good its individual statistics look.

Outcome Summary

Single-name funding-percentile continuation on LINK looks like amplified beta rather than an edge, so it should not be proposed again; a fair restatement would be a multi-name panel at fixed pre-registered thresholds, and sizing (notional_frac) should never be an optimizer free parameter.

The backtest review sent it on to optimization, but the analyst abandoned it after the hard holdout gate failed: holdout Sharpe was -0.69, 2.1 standard errors below OOS and not consistent with it. The failure was classified as overfit. The run also failed the programme FDR test, missed its own pre-registered drawdown and PF predictions, and its edge weakened as the funding threshold got more extreme, which contradicts the premise.

Go long the LINKUSDT Binance USD-M perpetual for a fixed 5 days (30 4H bars) whenever the trailing-72h mean funding rate is above the 80th percentile of its own trailing 90 days, on the view that crowded leveraged demand keeps pushing price up.

The backtest with the pre-registered settings produced 136 long trades over about 6.2 years: +106% total return, Sharpe 0.60 (CI -0.22 to 1.32), average trade +1.90% of notional, profit factor 1.24, win rate 44.9%, max drawdown 44.4%. The optimized configuration (entry_percentile ~42, notional_frac 0.745) was liquidated on 2024-10-09 for a -100% return, and the holdout Sharpe was -0.69 on 64 trades against a walk-forward OOS Sharpe of 1.42.

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