Strategies

ADA Spot Capitulation-Liquidity Bounce: Buy After a 24h Return in the Bottom 2% of Its Own Trailing 90-Day Distribution, Exit on a Fixed 24h Clock (ADAUSDT.BINANCE_SPOT 1H, long-only, no leverage, flat ~93% of the time, pure OHLCV)

Outcome: Abandoned

AdaSpotCapitulationTailBounce24h

Outcome Summary

AdaSpotCapitulationTailBounce24h tested a capitulation-liquidity bounce on ADA spot: buy after an extreme 24h selloff relative to ADA's own 90-day distribution and exit after a fixed 24h. Across two iterations it showed a strong in-sample edge (PF 1.63-1.90, 1.46-2.37% per trade, beta ~0.03) and clean walk-forward statistics (OOS Sharpe 0.96, PBO 0.32). The 20-month holdout, however, had essentially zero edge (26 trades, -0.04% per trade, holdout Sharpe 0.016), and performance faded year by year into 2026, so the analyst abandoned it as no_edge (edge decay) and noted it also failed deflated Sharpe and programme FDR. The analyst also flagged suspected, unverified MTM daily-return artifacts (a terminal-day -11%/-19% print and paired -X/+X days) for engine review.

Strategy report

Hypothesis

A single-leg, long-only, unleveraged, pure-OHLCV tail-event trade on ADAUSDT.BINANCE_SPOT with 1H bars. Mechanism, with instrument and timeframe removed: 'after an extreme short-horizon selloff relative to the asset's own recent distribution, supply liquidity to the forced or impatient sellers and exit on a fixed clock set at the peak of the post-event forward return'. This is the same mechanism family as the earlier DOGEUSDT.BYBIT immediacy-premium run (6h trigger, 12h clock). That run measured PF 1.39 with 8.9% drawdown and was risk_rejected, not falsified. This version changes the trigger window to 24h, the clock to 24h (the measured forward-return peak) and the venue to BINANCE_SPOT (8+ years of history, correct no-leverage venue). Point-in-time pre-study (trailing-90d 2% quantile shifted one bar, non-overlapping events, 0.20% round trip deducted): the net mean per trade is positive on 11 of 12 spot majors. On ADA: 218 trades (~26/yr), net +1.49% per trade, win rate 0.64, PF 1.74, t=2.9, calendar Sharpe ~0.97, 8 of 9 years positive. All 20 cells of the quantile (1/2/3/5%) x clock (12-36h) grid are positive, with PF 1.39-1.98. Flat about 93% of the time.

Iteration-2 fix for 'Sharpe -2.54 disagrees in sign with total_return +223%'. I reproduced it locally on the full history (Sharpe -2.71, +256%). The daily MTM series ended at -100% on 2020-03-13. The ADA holding was valued correctly (verified per bar). The cause was the post-hoc square-root market-impact cost: orders at 95% of a compounded equity against thin 2018-2020 ADA hourly volume ran up $409k of impact against $264k capacity, and folding that cost into the MTM curve bankrupted it. The only change is a liquidity cap on order size: qty <= 5% x the trailing 24-bar mean volume. Signal, entry trigger and 24h exit clock are unchanged. Local full-history rerun with the cap: total_return +488%, MTM Sharpe +0.33 (same sign), max DD 31%, PF 1.63, 215 trades, avg trade +1.48% (above the 0.25% spot floor). Every year is positive from 2018 to 2025; 2026 year-to-date is -21%. The large daily swings that remain (e.g. 2022-05-12) are real price moves during the LUNA crash, not accounting artefacts.

Failure code: no_edge (edge decay), taken from the first measured report. That report cleared the fee floor at 1.46%/trade against the 0.25% spot floor, with PF 1.63 but MTM Sharpe 0.33 and a CI straddling 0, so this is NOT fee_edge. It is not overfit in the PBO sense either: CPCV OOS is 0.96 and PBO 0.32. The ADA capitulation-bounce edge was real in 2018-2024 (134 in-sample trades, realized mean 2.27%, t = 3.25, PF 2.11) but is absent out of sample. The 20-month holdout has 26 trades with realized mean -0.04%/trade, PF 0.98, t = -0.04, and 2026 YTD is PF 0.62. The holdout 'pass' is an MTM Sharpe of 0.016, zero edge in substance. A waiver citing consistent_with_oos (z = -1.22) is not credible: that consistency reflects the noise of a small holdout sample, not forward evidence. It also sits alongside DSR 0.70 against an expected-max Sharpe of 0.40 over 15 effective trials, a bootstrap CI straddling 0, a programme-FDR failure, the missed pre-registered drawdown (21.3% vs 20%), and an MTM Sharpe of 0.59, below the 1.0 promote floor (the L165 flat-duty-cycle shape; the DOGE sibling was risk_rejected at Sharpe 0.48). Iterating would only re-tune the same decayed mechanism (L126). For the Research Lead: the pre-study found this mechanism positive on 11 of 12 spot majors. The only form worth re-proposing is a pooled multi-name panel (about 250+ trades a year), which would give enough holdout trades to test whether the post-2024 decay is ADA-specific or market-wide. Separately, a suspected engine artifact, not verified: a terminal-day MTM print of -11% (-19% in the base run) while the realized equity is flat and no position is open, plus recurring paired -X/+X daily prints. This should go to engine review.

Implementation

Long-only, unleveraged ADAUSDT spot on 1H bars. Buys when the trailing 24h return falls below the 2% quantile of its own trailing 90-day (2160-bar) distribution. The quantile is taken from prior bars only. Exits on a fixed 24h clock. Size is 95% of equity, capped at 5% of the trailing 24-bar mean hourly volume so that market impact stays bounded when volume is thin.

Verification Results

Optionally store (ts, close) and pick the base close as of bar.ts_init - 24h, or skip the signal when the span between _closes[0] and the current bar is more than 24h plus one bar.

The return window is 24 BARS, not 24 wall-clock hours. Missing 1H bars (exchange outages, catalog gaps) mean _closes[0] can be more than 24h old. Bars with close<=0 are skipped, which shifts the window. The effect on BINANCE_SPOT 1H is small.

Before paper, confirm that the live warmup replay feeds calculate_signal with at least 2160 bars.

The quantile needs min_dist_samples=1440 bars (60 days) of rolling state that lives inside calculate_signal. In backtest this is only a warm-up cost. In paper/live, a restart gives no entries for 60 days unless the base template's warmup replay drives calculate_signal.

No change is required for correctness. If the Risk Officer objects to the drawdown, prefer a lower equity_fraction over a stop the hypothesis does not name.

Size is 95% of equity with no per-trade stop, so a 24h hold carries the full move on nearly the whole account (sandbox largest loss $21.6k, DD 31.6%). This matches the hypothesis, which specifies no stop, and the design is flat by default with one entry per event rather than an always-in flip book. Drawdown will still be judged against the pre-registered floor downstream.

The sandbox window (834 days, mostly 2024-2026) is much weaker than the pre-study: PF 1.02, +3.1%, avg trade +0.66% of notional over 51 trades. Folded square-root impact ($13.5k) takes about 81% of gross, which fails the economics soft floor of impact <= 50% of gross. Per trade that is only about 0.13% on about $100k notional; the 5% participation cap does not bind on recent ADA volume, so this is not a sizing bug. The problem is that gross edge in this window is thin. The MTM Sharpe (-0.23) disagrees in sign with the +3.1% return because impact is folded into the MTM curve but not into the equity curve, and the metrics block already flags metrics_reliable=false. The regime split is lopsided: calm terciles lost 38.5% and stressed terciles made 45.3%. The edge may exist only in high-vol regimes, and a 2% quantile of a calm-period distribution may catch shallow 'tail' moves that do not bounce. What to check first in the full-history backtest: (1) the per-year split, especially 2024-2026 against 2018-2023; (2) impact as a share of gross over the full span; (3) whether a vol-regime filter or an absolute-return floor on the trigger belongs in the pre-registered design. Ledger magnitude check: 0.66% x 51 = 34% implied against 3.1% actual, about 11x, inside the ~20x tolerance. The signs of PF and return agree.

Backtest Review

Sharpe: 0.33
Total return: 480.86%
Max drawdown: 30.90%
Trades: 210
Win rate: 64.3%
Profit factor: 1.63

Trades match the mechanism: 210 trades, all long, each entered after a bottom-2% 24h selloff and exited on the fixed 24h clock. That is about 26 a year, and exposure is 7.2%, both matching the pre-registered design.

The trade ledger shows +0.97% per trade (reported avg_trade_return_pct is 1.46%), well above the 0.25% spot floor. PF is 1.63, win rate 0.64, fees are 6% of gross and impact 19% of gross.

The result does not depend on a few big trades: without the top 5 it is still +0.57% per trade and +$174k. Mean per-trade return is positive in 8 of 9 entry years (2018-2025).

It is not market beta: beta is 0.03 and correlation 0.11. QA's concern is only partly borne out, because the edge is present in 2018-2024 and not only in high-volatility periods.

The edge is weaker than pre-registered: t-stat about 1.9 against 2.9, and 0.97% against 1.49% per trade. It is fading recently: 2025 made +0.49% per trade and 2026 lost 1.89% per trade over 10 trades.

The mark-to-market Sharpe is 0.33 (CI -0.33 to 1.01), below the advisory bar of 1.0. It is distorted by pairs of daily returns that cancel exactly (-17% to -20% one day, then +21% to +25% the next) and by an unpaired -19.41% on the final day. The mechanism is not verified; flag it for engine review.

Max drawdown is 30.9% with a 649-day duration, partly caused by those same daily prints.

A book that is flat about 93% of the time has its calendar Sharpe capped (L165), so a later promote will be hard.

Analysis

Mechanism matches the hypothesis: 160 long-only entries after 24h returns in the bottom ~1.3% tail, fixed ~23h clock, flat ~94% of the time.

Per-trade economics are far above cost: realized mean 1.90% (reported avg_trade_return_pct 2.37%) against the 0.25% BINANCE_SPOT floor; full-sample realized t = 3.06, PF 1.91, win rate 0.66.

Not beta: beta 0.03, correlation 0.16. The drawdown (21%) is far below ADA buy-and-hold drawdown (>80%).

CPCV walk-forward is clean: OOS 0.96, PBO 0.32, not overfit; 8 of 9 years positive; no single year dominates (2021 carries 24%); sensitivity is smooth apart from the equity_fraction>1 clamp and a short-hold edge on hold_hours.

Forward edge is absent. On the realized trade ledger, the holdout (2024-12-13 to 2026-08-13) has 26 trades with mean -0.04%/trade, t = -0.04, PF 0.98, win rate 0.46. The HARD holdout gate 'passed' only on an MTM Sharpe of 0.016 (ratio 0.017), which is zero edge in substance.

The decay is monotone at the end of the sample: realized mean per trade was 2.46% in 2024, 0.71% in 2025 (PF 1.41) and -1.35% in 2026 (PF 0.62, 8 trades).

Deflation fails: DSR 0.70 against an expected-max Sharpe of 0.40 over 15 effective trials (82 raw). The bootstrap Sharpe CI of -0.11 to 1.25 straddles 0, and the strategy does not survive programme FDR (p = 0.30; 5 of 309 kept).

Pre-registered drawdown missed: 21.3% against the declared 20%. 5 of 6 pre-registered checks were met.

MTM Sharpe of 0.59 is below the Sharpe 1.0 promote floor. This is the flat ~93%-of-the-time shape from L165, the same shape the DOGE immediacy-premium sibling had when it was risk_rejected.

Suspected engine artifact, not verified: the optimized daily series ends with a -11.0% print on 2026-08-14 (-19.4% in the base run) while the realized equity curve is flat at $426,251 with no open position. There are also recurring paired -X/+X days on entry/exit days. These distort the MTM Sharpe and should go to engine review. They do not drive this verdict, which rests on the realized ledger.

not kept

kept

Backtest gate: - metrics unreliable: Incoherent: Sharpe -2.54 disagrees in sign with total_return 223.25% — the Sharpe is computed from the mtm_equity_curve series while the return comes from the equity curve. Judge on the equity curve.

Outcome Summary

A single-name tail-event trade yields too few holdout trades to tell decay from noise, so the analyst recommended re-proposing it only as a pooled multi-name spot panel (about 250+ trades a year), which could also show whether the post-2024 decay is specific to ADA.

The analyst passed the backtest review (optimize) but abandoned the strategy after optimization with failure code no_edge (edge decay): the edge was real in 2018-2024 but absent out of sample, with realized per-trade mean falling from 2.46% in 2024 to 0.71% in 2025 and -1.35% in 2026. It also failed deflated Sharpe (0.70 vs 0.95), programme FDR (p=0.30, 5 of 309 kept), the 1.0 Sharpe promote floor and its pre-registered 20% drawdown limit.

A long-only, unleveraged ADAUSDT.BINANCE_SPOT 1H strategy that buys after the trailing 24h return falls into the bottom ~2% of its own trailing 90-day distribution, supplying liquidity to capitulating sellers, and exits on a fixed ~24h clock, staying flat about 93% of the time.

The base backtest (2018-04 to 2026-08) made +480.9% over 210 trades with PF 1.63, win rate 0.64, 1.46% avg trade return, MTM Sharpe 0.33 and 30.9% max drawdown; the optimized version made +326.3% over 160 trades with PF 1.90, 2.37% avg trade return, Sharpe 0.59, 21.3% drawdown, and CPCV walk-forward OOS Sharpe 0.96 with PBO 0.32. The 608-day holdout nominally passed at Sharpe 0.016 (ratio 0.017 vs 0.7 floor), and on the realized ledger its 26 trades averaged -0.04% per trade with PF 0.98.

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