AdaFourHourNBarCumulativeReturnZScoreMomentumLong

ADA 4H N-Bar Cumulative Return Threshold Long (Multi-Bar Momentum Accumulation)

A long-only, single-instrument, single-timeframe strategy on ADAUSDT.BINANCE 4H bars (the most-reliably-fresh bar type). Implements a NEW MECHANISM not yet tried in the portfolio: N-BAR CUMULATIVE RETURN THRESHOLD. The mechanism: compute the cumulative log return over the last 6 bars (24h on 4H = 1 day of momentum), standardize it as a z-score against the trailing 100-bar distribution of 6-bar cumulative returns, and enter long when this z-score exceeds 1.0 (cumulative 6-bar return is in the top ~15% of historical 6-bar returns) AND the current bar closes bullishly AND in confirmed uptrend. Captures SUSTAINED multi-bar momentum as a single composite signal — distinct from single-bar return z-score (which only sees one bar's return), distinct from sequence-based HH counting (which uses bar-extreme sequences, not return magnitudes), distinct from autocorrelation (which uses correlation, not magnitude). This is the TSMOM (Moskowitz-Ooi-Pedersen 2012) signal applied at intraday resolution: when the past day's cumulative return is unusually strong AND in uptrend, momentum tends to extend over 1-3 days. Three-filter simplicity (cumulative z-score + bullish bar + trend filter) per analyst's recurring guidance.

Implements the hypothesis literally as intraday TSMOM with a clean three-filter design (cumulative-return z-score + bullish bar + trend). The mechanism is a NEW signal class for the portfolio: the per-bar observation is a MULTI-bar statistic (6-bar cumulative log return) which is then z-scored against its own trailing distribution — structurally different from a single-bar return z-score, HH-sequence counting (bar extremes), or autocorrelation (correlation rather than magnitude). Log returns are used because they are additive, so the 6-bar cumulative return is simply logclose[t]-logclose[t-6], avoiding compounding error. The signal is continuous and varying (the z-score returned every bar) while the binary confirmations live in should_enter via stored flags, satisfying the 'continuous signal, gated entry' rule and giving the verifier a non-frozen signal. Long-only with a trend gate matches the momentum-extension bias and avoids buying downtrends; the 4% stop / 8% take-profit give per-trade economics well above the ~0.10% round-trip futures fee, and the 18-bar time stop bounds dead trades. Kept single-instrument on ADAUSDT.BINANCE 4H (most-reliably-fresh path, no supplementary data, no multi-instrument staleness); leverage=1.0 because sizing uses no margin amplification, avoiding the leverage_set_but_unused gate.

Genuine-but-too-weak momentum edge that fails multiple-testing deflation and the forward holdout -- exactly the outcome flagged at pre-optimization. The optimizer DID fix the cost problems (impact 22.7%->6.35%, capacity $1.9M->$24.8M) and produced a config positive in 6 of 7 years (PF 1.49, sharpe_ci_low positive 0.42, PBO 0.4974), but the decisive gates reject it: deflated_sharpe 0.041 (is_significant FALSE) with the optimized Sharpe 2.07 BELOW the expected-max luck bar 3.40 over 225 trials -- after honest multiple-testing correction the edge is indistinguishable from best-of-N noise; the HOLDOUT FAILED with a deeply NEGATIVE Sharpe -2.42 (the recent one-shot forward window loses money badly); walk-forward is_overfitted=TRUE (avg IS 3.32 -> avg OOS 0.822, one OOS window -1.02); and sensitivity FAILED with 3 cliffs on the CORE signal params (entry_z, cum_bars, trend_sma). The 6/7-positive-years and positive CI are in-sample/walk-forward best-of-225 artifacts; the forward test and DSR both fail, and the edge is decaying (2026 -4.0%, rolling Sharpe negative through 2026). Not iterate (attempt 1 of 2): the failure is signal weakness (DSR 0.041, below the luck bar) + a negative holdout + cliffs on the core params, not a tunable parameter region -- a second optimization re-finds the same sub-luck-bar / negative-holdout edge, and the cost fixes don't matter if the signal can't deflate. Not revise_hypothesis: TSMOM on a single crypto major is the premise and it is marginal everywhere tried. FAILURE PATTERN: time-series/cumulative-return momentum on a single crypto major produces a genuine ~0.5 base Sharpe that, after best-of-225 selection, fails deflated Sharpe (0.041, below the expected-max luck bar 3.40), fails the holdout with a deeply negative Sharpe (-2.42), and is_overfitted=TRUE with core-param cliffs -- the same DSR-fail signature as the cross-sectional momentum sibling (DSR 0.05) and the ADA OBV-divergence (DSR 0.0007). The momentum factor on a single major is too weak to deflate regardless of fixing impact/capacity; a positive walk-forward with a negative holdout and DSR 0.04 is an edge fit to historical regimes that is decaying out-of-sample.

Long-only single-instrument momentum strategy on ADAUSDT.BINANCE USD-M futures 4H bars, OHLCV-only. Each bar it computes the 6-bar cumulative log return (~1 day of momentum) and standardizes it as a z-score against the trailing 100-bar distribution of 6-bar cumulative returns. It enters long when (1) that z-score >= 1.0 (the past day's drift is in the top ~15% of its own history), (2) the current bar closes bullishly (close > open), and (3) the long-term trend filter holds (close > SMA(50)). It exits on a 4% stop-loss, 8% take-profit, an 18-bar (~3 day) time stop, the momentum z fading back to <= 0, or the trend breaking. calculate_signal returns the cumulative-return z-score continuously every bar; the bullish-bar and trend gates are stored flags ANDed in should_enter. Sizes 30% of equity per trade, capped by 1.5% risk over the 4% stop.

Genuine TSMOM momentum payoff: win_rate 0.368 with avg_win $2,120 vs avg_loss $1,100 (1.9x), tail_ratio 1.73, skew 1.84 -- big-winner momentum extension, not a fade

Positive in 5 of 7 years spread across regimes (2020-2024, incl. the 2022 bear) -- real breadth, not front-loaded; ample sample (408 trades), PF 1.12, total_return +44%

Sharpe 0.47 with sharpe_ci_low -0.86 (CI straddles 0) -- weak, not yet significant; primary DSR-fail risk after selection

impact_cost_pct 22.7% (near threshold) and capacity_usd only $1.9M (toy scale) -- largely turnover-driven (turnover 129.8), so partly tunable but a real promotion concern

Recent decay: 2025 -1.2%, 2026 -4.5% (recent rolling Sharpe deeply negative)

deployable scale

~0.5+ (pre-opt)

~62 WF floor

5 of 7

majority

1.9x (avg_win/avg_loss)

> 1 for momentum

Optimizer fixed the cost problems flagged at pre-opt: impact 22.7%->6.35%, capacity $1.9M->$24.8M; optimized backtest positive in 6 of 7 years, PF 1.49, max_dd 9.8%

sharpe_ci_low positive (0.42), PBO 0.4974 (just under 0.5) -- the in-sample/walk-forward fit is not egregiously overfit

Fails deflated Sharpe: DSR 0.041 (<<0.95), is_significant FALSE; optimized Sharpe 2.07 BELOW the expected-max luck bar 3.40 over 225 trials -- noise-indistinguishable after multiple-testing correction

Holdout FAILED with a deeply NEGATIVE Sharpe -2.42 (ratio -2.944) -- the recent one-shot out-of-sample window loses money badly

Walk-forward is_overfitted TRUE (avg IS 3.32 -> avg OOS 0.822, one OOS window -1.02); sensitivity FAILED with 3 cliffs on the core signal params (entry_z, cum_bars, trend_sma)

Recent decay: 2026 -4.0%, rolling Sharpe deeply negative through 2026

> 3.40 (expected-max)

AdaFourHourNBarCumulativeReturnZScoreMomentumLong was the session's strongest momentum attempt — a clean TSMOM signal with a genuine, non-inverted payoff (avg_win 1.9× avg_loss, big-winner skew) and real multi-year breadth, positive in 5 of 7 years including the 2022 bear. It cleared the review gate, and optimization did exactly what was asked: it fixed the turnover-driven cost problems (impact 22.7%→6.35%, capacity $1.9M→$24.8M) and produced a config positive in 6 of 7 years with a positive Sharpe CI and PBO just under 0.5. But the decisive forward gates rejected it — the optimized Sharpe of 2.07 fell below the 3.40 luck bar (DSR 0.041), the once-tested holdout lost money badly at −2.42, the walk-forward was overfitted, and all three core signal parameters were cliffs. The analyst abandoned it on its first iteration: fixing cost can't rescue a signal that won't deflate, and this joins the cross-sectional-momentum and OBV-divergence results as confirmation that the momentum factor on a single crypto major is genuine but too marginal to survive honest multiple-testing correction and a forward holdout.

Time-series/cumulative-return momentum on a single crypto major produces a genuine ~0.5 base Sharpe that, after best-of-225 selection, fails deflated Sharpe (0.041, below the 3.40 luck bar), fails the holdout (−2.42), and is overfitted with core-param cliffs — the same DSR-fail signature as the cross-sectional momentum (DSR 0.05) and ADA OBV-divergence (DSR 0.0007) siblings — so the momentum factor on one major is too weak to deflate regardless of fixing impact/capacity, and a positive walk-forward with a negative holdout is an edge fit to historical regimes, not a real one.

The backtest-review gate passed it to optimization ('optimize'), but the analyst abandoned it post-optimization: despite a positive CI and the cost fixes, the optimized Sharpe 2.07 sat BELOW the expected-max luck bar of 3.40 over 225 trials (DSR 0.041, is_significant=false) so the edge is indistinguishable from best-of-N noise; the once-tested holdout failed badly (Sharpe −2.42); walk-forward was overfitted; and the defining entry_z/cum_bars/trend_sma parameters were all sensitivity cliffs — a signal-weakness failure, not a tunable region, with the edge decaying out-of-sample (2026 −4.0%).

A long-only, OHLCV-only TSMOM (time-series momentum) strategy on ADAUSDT.BINANCE 4H bars implementing a new mechanism — the 6-bar (~1-day) cumulative log return, z-scored against its trailing 100-bar distribution — entering long when that z-score ≥ 1.0 (top ~15% of historical 6-bar drift) AND the bar closes bullishly AND price is above its trend SMA, exiting on momentum fade, stop, take-profit, trend break, or time stop.

The raw backtest showed a genuine, non-inverted momentum payoff — 408 trades, win_rate 0.368 but avg_win $2,120 vs avg_loss $1,100 (1.9×), tail_ratio 1.73, skew 1.84, profit_factor 1.12, total_return +44%, positive in 5 of 7 years — though weak (Sharpe 0.47, sharpe_ci_low −0.86) and turnover-heavy (impact 22.7%, capacity $1.9M). Optimization fixed the cost problems (impact 6.35%, capacity $24.8M) and lifted it to Sharpe 2.07 / PF 1.49 / +88% with a positive Sharpe CI (0.42) and PBO 0.4974, but failed the forward gates: DSR 0.041, holdout Sharpe −2.42 (ratio −2.944), walk-forward is_overfitted=TRUE (IS 3.32 → OOS 0.822), and 3 sensitivity cliffs on the core signal params.