Strategies

SOL Rejected-New-High Probe Continuation: Go Long for a Fixed 36h After an Hourly Bar Pierces the Prior 20-Day High but CLOSES Back Below It (SOLUSDT.BINANCE USD-M 1H, long-only, 1 structural parameter + clock, flat ~89% of the time, pure OHLCV)

Outcome: Abandoned

SolRejectedNewHighProbeLong

Outcome Summary

The hypothesis bet on continuation after SOL's rejected probes of a multi-week high, based on a pre-study showing the effect on 9 of 10 perps. After one iteration, the backtest cleared costs easily (2.72% per trade over 145 trades, beta 0.23), and the backtest review sent it to optimization. The optimized 54h/28d version reached a walk-forward OOS Sharpe of 1.17 but lost on the 2025-2026 holdout (Sharpe -1.09 on 14 trades), and PBO came in at 0.587. The analyst abandoned it as overfit, because the returns were concentrated in 2021 and the unoptimized parameters also decayed in 2025-2026.

Strategy report

Hypothesis

Single-leg, single-instrument, long-only EVENT trade on the SOLUSDT.BINANCE USD-M perpetual, 1-HOUR bars, pure OHLCV. It is the evidence-driven INVERSE of a family the factory has killed at least 15 times: the 'failed-breakout / liquidity-sweep / upthrust FADE' (BNB False-Breakout Trap Reversal, XRP Failed-Breakout Reversal, SOL Liquidity-Sweep Reversal, SOL COIN-M Failed-Breakout Trap Fade, AVAX/DOGE/BTC sweep fades, ETH COIN-M failed-breakout exhaustion - all abandoned). Those runs shorted the bar that probed a multi-day high and closed back inside the range. If that fade reliably loses, the other side of it - buying the rejected upside probe - is where the money went. A point-in-time pre-study (run 2026-09-29 on Binance USD-M 1H bars, 2020-2026) confirms it: event = first hourly bar of a UTC day whose HIGH exceeds the max high of the prior 20 COMPLETED UTC days while its CLOSE is back below that level; enter at that bar's close; hold a fixed clock; non-overlapping. Forward mean return after the event beats the unconditional drift on 9 of 10 liquid USD-M perps (SOL, ETH, BNB, AVAX, XRP, DOGE, ADA, TRX, LINK; LTC is the exception). On SOL the per-hour-after-trigger curve peaks at 36h (per L166): 12h +0.64%, 24h +1.49%, 30h +1.96%, 36h +2.54%, 42h +2.25%, 60h +2.75% mean (unconditional 36h drift +0.51%). At the 36h clock: 157 trades, win rate 58%, calendar Sharpe 1.24 at 1x, 11% time in market. It is robust to the lookback: calendar Sharpe 1.19 / 1.24 / 1.39 / 1.22 at N = 10 / 20 / 30 / 55 days. Honest caveats from the same study: 2021 carries ~55-60% of the summed return and 2025 was negative (-12%), so the pre-registered OOS prediction is well below the in-sample 1.24. The mechanism stated without instrument or timeframe - 'buy a rejected probe of a multi-week high and hold it for about a day and a half' - does not match anything abandoned. The abandoned corpus holds only the opposite sign (fade the probe) and close-confirmed Donchian breakouts, which enter on a CLOSE above the level and at a higher price. SOL, not BTC, keeps it off the 44-62% BTC concentration. It stays on BINANCE USD-M rather than HL (L168: full multi-year history, 157 lifetime trades is at the ~150 floor, more on HL would be impossible). No supplementary data, no multi-leg book (L137), no signal-on-A/trade-B (L143), not a port (L167), and it has a declared flat state (L150).

This implements the pre-study's event definition exactly: prior 20 completed UTC days, the first qualifying bar of the day, entry at that close, a 36h clock and non-overlapping trades. The one structural parameter is lookback_days, plus hold_hours for the clock. min_event_pct is 0 by default (any rejected probe qualifies). The trade is long-only but stays on BINANCE USD-M futures, as the hypothesis specifies for full history and parity with the pre-study; the analyst may note that it could run on BINANCE_SPOT. Per-bar cost is O(20), because the daily-high deque is capped. The expected trade count is about 157 over 2020-2026 (about 25/yr), so the sandbox window should still see several trades. Caveat from the hypothesis itself: returns are concentrated in 2021.

Failure code: overfit (holdout failure plus PBO). Both hard validity gates failed. On 14 holdout trades from 2025-06 to 2026-08, Sharpe was -1.09 (z = -1.95 against WF-OOS 1.17, consistent_with_oos = false), and PBO is 0.587. This is not fee_edge: the first measured report showed avg_trade_return_pct of 2.72% against a 0.15% floor, and the optimized report 4.00%. It is not beta either: drawdown was 28-41% with beta 0.13-0.23, far below SOL buy-and-hold. The rejected-probe continuation edge on SOL is real in-sample, but it is a 2021/2023 bull-market effect. 2021 carries 65% of the profit, only 3 of 7 years are positive, and the UNOPTIMIZED 36h/20d base also lost money in 2025 (-10.8%) and 2026 (-24.9%). The decay therefore is not an optimizer artifact, and no retune of the clock or lookback can recover a holdout that is negative for the base parameters too. Note for the Research Lead: the single-instrument SOL form is dead. The pre-study reports that the effect holds on 9 of 10 perps, but that was measured over the same 2020-2026 window that carries the 2021 concentration. A panel restatement would only be worth proposing if its per-name 2025-2026 forward means are positive BEFORE coding. Do not re-propose SOL-only or other single-name versions of 'buy the rejected N-day-high probe'.

Implementation

Long-only 1H event trade on SOLUSDT.BINANCE USD-M. The level is the max high of the prior 20 COMPLETED UTC days, taken from its own daily-high deque with bars assigned to the UTC day of their open. The first hourly bar of a UTC day whose high exceeds that level and whose close is back below it triggers a market buy at the bar close. The position is held on a fixed 36h calendar clock anchored on position.ts_opened, so it survives restarts, then closed. Only one event is counted per UTC day. Entries happen only when flat, so events never overlap. The signal is continuous every bar: min(high-level, level-close)/level*100 (%), positive only on a rejected probe. Sizing is 95% of equity x leverage (1.0) / price. There is no stop, no supplementary data and no extra legs.

Verification Results

Keep the mechanism as it is. The analyst should read the worst-trade loss and the drawdown against the pre-registered max_drawdown_pct. If needed, add a catastrophe stop, for example a bracket at -12%, as an optional parameter rather than changing the tested rule.

There is no stop. Each event deploys 95% of equity at 1x for 36h, so the worst per-trade loss equals SOL's worst 36h drawdown, which can be 20-30% in a crash. The sandbox's largest loss was -$15.3k, about 13-15% of equity. The hypothesis pre-registers a pure clock exit, so this is faithful, and it is not an always-in compounding design: the book is flat about 89% of the time, sizing is fixed and there is no dynamic scalar. Worst-case gross = 0.95 x 1.0 x 1 leg = 0.95x equity, which is within the venue cap. Per-trade risk still exceeds the 1-2% guideline (L148).

Optional: when day jumps by more than 1, push placeholder entries (or skip) so that the deque tracks calendar days exactly.

The lookback counts completed UTC days that have bars, not calendar days. A whole missing day in the feed would silently stretch the 20-day window by one day. The first partial day after warm-up is also recorded with a partial high. Both effects are negligible on Binance USD-M 1H data.

No code change. The analyst should account for funding drag, or compare against a SOLUSDT.BINANCE_SPOT run.

The design is long-only at 1x leverage, which the venue-routing rule would place on BINANCE_SPOT. The hypothesis explicitly specifies BINANCE USD-M for full history and parity with the pre-study, so the choice is not a mismatch. However, the backtest does not fold funding (funding_folded=false), so the perp PnL here overstates a real long-perp position in positive-funding regimes.

The sandbox covers 834 days: PF 1.24, Sharpe 0.45 with a CI of -0.83 to 1.62, and 47 trades. The calm vol tercile lost 11% and the stressed tercile made 28%, so the edge is concentrated in high-vol episodes. The hypothesis says the same about 2021: it carries 55-60% of the return, and 2025 was negative. Funding is NOT folded (funding_folded=false, 0 funding events). A long SOL perp held 36h crosses about 4-5 funding stamps. In 2021-type bull regimes that can cost 0.1-0.4% per trade, which is a material share of the 0.82% avg trade. The analyst should check per-year returns, whether 2021 dominates, and whether the result survives an estimate of funding drag, or rerun it on BINANCE_SPOT, where there is no funding but the round-trip fee is 0.20%.

Backtest Review

Sharpe: 1.00
Total return: 1546.80%
Max drawdown: 40.63%
Trades: 145
Win rate: 57.2%
Profit factor: 1.54

The trades implement the mechanism as stated. All 145 trades are LONG, each holds exactly 1d 12h (the 36h clock), and entries are non-overlapping. The count of 145 is close to the pre-study's 157 events. There is 1 structural parameter plus the clock.

Sample: 145 trades over 5.9 years. That clears the ~100-trade bar in L178, and the holdout (last 20%) should hold about 25-30 trades, above the 10-trade minimum.

Fee edge: avg_trade_return_pct is 2.72% against the 0.15% floor for BINANCE USD-M, about 18x. The median trade is +1.12%. Commissions are 3.4% of gross. PF 1.54, win rate 0.57, Sharpe 1.00 with a bootstrap CI of 0.21 to 1.76 (does not straddle 0), PSR 0.85.

Not beta (L170). Beta is 0.23 and exposure 10.5%. The strategy's max drawdown of 40.6% is about 0.42x buy-and-hold SOL's drawdown of roughly 95% in 2021-22, below the 0.7x beta threshold.

QA's funding concern is not borne out. Funding IS folded in the full backtest (funding_folded=true, 635 events), unlike the sandbox. Total per-trade funding was -$80k against +$1.55M net PnL, about 5% of PnL. The edge survives funding.

Win-rate/PF (0.57/1.54) is not the falsified payoff shape in L142.

QA's concentration concern IS borne out. By exit year, summed pnl_pct is: 2020 -11.6% (2 trades); 2021 +190% (44); 2022 +9.6% (10); 2023 +98% (30); 2024 +60% (28); 2025 -5.5% (20); 2026 -5.6% (11). The last ~20 months average a negative per-trade return. That is roughly the holdout window, so holdout failure is a real risk. This is flagged but not predicted as a verdict, per the review rules.

Regime split: the calm vol tercile loses (-40%, Sharpe -0.26). Normal and stressed are strongly positive.

Drawdown is 40.6% (CI up to 81%) and took 753 days to recover. The book deploys at about 1x notional, so position_fraction is effectively the drawdown knob.

Suspected metrics artifact, mechanism not verified. daily_returns contains paired -X%/+Y% days that nearly cancel while no trade is open: 2021-09-16/17 and 2021-09-29/30 (both -9.517/+10.518), 2024-03-21/22 and 2024-04-03/04 (both -12.4151/+14.175), 2025-03-15/16 and 2025-03-28/29 (both -9.4503/+10.4366). There is also a -19.75% print on 2026-08-17 while the book is flat, although the equity curve stays constant at $1,646,795 over that period. These phantom days inflate the daily-return volatility, the Sharpe denominator and the 2026 annual figure (-24.9%, of which about -19.75 is phantom). Engine review of the daily_returns derivation is needed. It does not change the trade ledger.

Analysis

Per-trade economics are far above costs. The first backtest made 2.72% per trade (145 trades) and the optimized one 4.00% (100 trades), against a 0.15% futures floor. This is not a fee_edge death.

It is not beta (L170). The drawdown was 40.6% at base and 28.3% optimized, against a much deeper SOL buy-and-hold drawdown over 2021-2023. Beta is 0.13-0.23, time in market is about 11%, and alpha is 0.22-0.32.

The parameter surface is smooth. There are no cliffs, and the hold_hours by lookback_days heatmap is flat at 0.65-1.03. CPCV average OOS Sharpe is 1.17, and DSR is 0.975 against an expected maximum of 0.33 over about 14 effective trials.

The trades match the mechanism: long-only, one entry per UTC day on a rejected probe of the N-day high, a fixed clock exit, and zero unrealized P&L at the end.

HARD: the holdout failed. On 2025-06-10 to 2026-08-16 (432 days, 14 trades), Sharpe was -1.09 and the ratio was -0.93. z_vs_oos is -1.95 and consistent_with_oos is false. The edge did not appear on unseen data.

HARD: PBO is 0.587, above 0.5. The chosen parameters (hold 54h, lookback 28d) were the in-sample winner in only 27.8% of splits. The optimizer moved away from the pre-registered 36h / 20d towards a longer hold and lookback, which is fitted to the 2021 and 2023 rallies.

The decay is visible in both reports, not only in the holdout. The unoptimized base lost money in 2025 (-10.8%) and 2026 (-24.9%). The optimized version lost 14.3% in 2025 and 3.5% in 2026, and 2022 was flat to negative. Only 3 of 7 calendar years are positive (43%, below the 60% floor that is soaking to hard). 2021 carries 65% of the profit, which the hypothesis's own pre-study warned about (55-60%).

The regime profile is one-sided. The strategy loses in calm-vol terciles (Sharpe -0.23 optimized, -0.26 base) and all of its profit comes from normal and stressed regimes, which were dominated by the 2021 and 2023 SOL bull legs.

Pre-registered predictions missed 2 of 6: min_trades was 140 against 100 actual, and max_drawdown was 20% against 28.3% actual. The strategy also does not survive the programme FDR (p=0.025 against a threshold of 0.0015), and is_significant is false.

> 0 on >= 10 trades

Benjamini-Hochberg at q=0.10 over 329 programme candidates keeps 8. A candidate that does not survive here is not distinguishable from the programme's own noise, however good its individual statistics look.

Outcome Summary

A single-name 'buy the rejected N-day-high probe' edge can be a 2021/2023 bull-market effect. Its unoptimized base also lost money in 2025-2026, so a panel version should only be proposed if per-name 2025-2026 forward means are positive before coding.

The analyst abandoned it as overfit because two hard validity gates failed: the holdout was negative (ratio -0.93, z -1.95, not consistent with OOS) and PBO was 0.587, above 0.5. In addition, 2021 carried 65% of the profit, only 3 of 7 years were positive, and it did not survive the programme FDR.

Go long SOLUSDT (Binance USD-M, 1H) for a fixed 36h after the first hourly bar of a UTC day probes above the prior 20-day high but closes back below it. This is the opposite side of the failed-breakout fade the factory had abandoned many times.

The initial backtest made 145 long trades with a 2.72% average trade return, Sharpe 1.00, profit factor 1.54 and a 40.6% max drawdown. The optimized version (54h hold, 28-day lookback) made 100 trades at 4.00% per trade, Sharpe 1.12, with a 28.3% drawdown and a walk-forward OOS Sharpe of 1.17. The holdout (2025-06-10 to 2026-08-16, 14 trades) had a Sharpe of -1.09.

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