SolIntermediateSkipWindowTsmomLS

SOL Intermediate (Skip-Window) Time-Series Momentum, Long-Short (SOLUSDT.BINANCE USD-M, 4H Bars — Trend Measured OVER t-N..t-k EXCLUDING the Most Recent k Bars to Strip Short-Term Reversal Whipsaw, Volatility-Scaled Sizing, ATR-Trailing Exit, 3-Parameter)

A LONG-SHORT, single-instrument, pure-OHLCV directional strategy on SOLUSDT.BINANCE USD-M perpetual (4H bars). The mechanism is INTERMEDIATE / SKIP-WINDOW time-series momentum: the trend signal is the cumulative return over t-N..t-k EXCLUDING the most recent k bars, where short-term reversal lives and manufactures the sub-fee whipsaw that killed prior SOL/altcoin trend clones. Sign of the intermediate momentum sets direction with a deadband; inverse-volatility sizing; single-instrument, same-venue, no fades.

Implements the hypothesis literally: the skip window is the core mechanism, so the numerator is the cumulative LOG RETURN from t-N to t-k with the last k bars deliberately dropped from the measurement (they still update ATR, price and the trailing stop, they just do not vote on direction). Normalising by ATR%*sqrt(N-k) makes the score scale-free across SOL vol regimes, so the single deadband parameter has the same meaning in calm and stressed markets (per L5: no fixed absolute-return threshold). The signal is continuous and returned on every bar; the entry decision is a pure threshold in should_enter. Three tuned parameters (lookback, skip, deadband) plus ATR trail/period and risk fraction, no regime/volume/confluence conjunction stacked on top (per L25) so entries fire often enough for the sandbox window. Warmup is 49 bars, a tiny fraction of the 4H sandbox window (per L21). Futures venue BINANCE is required because the strategy goes short; leverage stays 1.0 and gross notional is capped independently of the equity path so losses cannot compound into position size (per L15). Sizing is quantised to SOLUSDT size precision (3 dp) and goes through the base template's min-notional guard.

Overfit single-name OHLCV TSMOM clone. IS Sharpe 2.96 collapses to OOS -0.099 (is_overfitted=true), holdout Sharpe -0.228 on 59 trades, PBO 0.51 > 0.5, DSR 0.0 / is_significant=false / fails programme FDR — the selected Sharpe is indistinguishable from best-of-N noise over 225 trials. The skip-window reframe added no robustness the 3-phase deflation could not strip; the edge is absent, not mis-tuned (lookback is itself a cliff parameter, no robust region to iterate toward). Matches the single-name daily/4H TSMOM-clone overfit family (prior SOL TSMOM/momentum-confluence clones all abandoned fee_edge/no_edge). Per the decision framework, overfit -> abandon (reframing the same mechanism onto another instrument would just overfit again).

Long-short intermediate (skip-window) time-series momentum on SOLUSDT.BINANCE USD-M perpetual, 4H bars. Every bar it computes a continuous momentum score = log(P[t-k]/P[t-N]) divided by the ATR-implied move over the same (N-k)-bar horizon, i.e. the trend is measured over t-N..t-k and the MOST RECENT k bars are excluded so short-term reversal noise never enters the trend estimate. Score above +deadband goes long, below -deadband goes short. Exit on an intermediate-momentum sign flip or an ATR trailing stop off the favourable extreme since entry. Sizing is inverse-volatility: a fixed fraction of equity risked over the ATR stop distance, hard-capped at a fraction of equity notional. Defaults: N=30 bars (5 days), k=3 bars (12h), deadband 0.6 sigma, 3x ATR(14) trail, 2% risk per trade.

If the intent was to set the framework warmup, confirm the base actually consults min_bars_required; otherwise rename to an internal helper (e.g. _min_bars) to avoid the shadow warning.

min_bars_required(self) overrides/shadows a base framework method; static analysis flags that the base may rely on its own warmup implementation and ignore this override. This is harmless here because calculate_signal has its own internal warmup guard (returns 0.0 until len(_closes) > lookback), so signals are never emitted on insufficient data regardless of which warmup path the base uses.

Single-name pure-OHLCV directional trend follower on SOL, a family that has repeatedly died at PF~1.0 after optimization. Sandbox is positive (Sharpe 0.64, PF 1.27, avg_trade 1.17%) but the Sharpe CI straddles zero (low -0.15) and win rate is 43% with gains concentrated in the normal/stressed vol terciles (calm tercile only +5.5%, Sharpe 0.24). The skip-window construction is a genuinely new mechanism (per L81, not a re-parameterization of an abandoned OHLCV mechanic), so it deserves a real backtest, but the analyst should confirm the intermediate-momentum edge survives full-history costs across all vol regimes rather than living only in trending regimes.

Decisive sample: 362 trades over a 6-year (2183-day) window, metrics_reliable=true.

Per-trade edge is genuinely above costs: avg_trade_return_pct 1.17% vs the 0.15% Binance USD-M floor; commission only 3.3% of gross, impact 7.1% — this is NOT fee-fragility.

PF 1.237 sits clearly ABOVE the [0.85,1.10] fee/impact drag band, positive Sharpe 0.643, positive alpha 0.10, DD 22.2% (<60%), recovery_factor 4.17.

Trades implement the hypothesis: balanced long (171) / short (191), 362 signaled = 362 submitted, ATR-trailing exit, inverse-vol sizing. No mechanism mismatch.

Skip-window (intermediate, reversal-stripped) TSMOM is a genuinely new construction per L81/QA, not a re-parameterization of an abandoned OHLCV mechanic — earns a real test.

Profitable across all three vol terciles (calm +5.5%, normal +32%, stressed +45%; only_one_regime=false).

Classic single-name OHLCV trend family (L82/L93) that repeatedly dies overfit in the 3-phase deflation — this base Sharpe of 0.64 is below the ADA survivor tier (Sharpe 1.6/PF 2.19).

Edge is decaying by calendar year: 2021-2023 strongly positive, but 2024 -1.5%, 2025 -7.2%, 2026 +1.5% — recent regimes barely contribute, and holdout is the last ~15 months.

Sharpe 95% CI [-0.17, 1.45] straddles zero on the low end; win rate 43%; calm-tercile Sharpe only 0.24.

The holdout window falls squarely in the weak 2025-2026 period, so the SOFT deflation/holdout gates may be hard to clear.

Decisive sample (403 trades, 2186-day window, metrics_reliable=true) and a genuine above-cost per-trade edge (avg_trade_return_pct 0.80% vs the 0.15% Binance USD-M floor; commission 3.99% of gross, impact 9.4%) — the failure is NOT fee-fragility.

Sensitivity grid is mostly smooth for most parameters (only lookback flagged as a cliff).

Overfit (HARD): in-sample Sharpe 2.96 collapses to out-of-sample -0.099; walk-forward is_overfitted=true. Per-window OOS Sharpes are -0.40 / +0.10 / 0.0 — no edge generalizes.

Holdout FAILED (HARD): holdout Sharpe -0.228 over 59 trades — a negative forward edge on the last ~15 months of unseen data.

PBO 0.512 > 0.5 (HARD): parameter selection is more likely than not overfitting.

Robustness collapse: deflated_sharpe 0.0, PSR 0.85, is_significant=false, fails programme FDR (keeps 4 of 269). Selected Sharpe is indistinguishable from best-of-N noise over 225 trials; expected-max Sharpe 2.75 exceeds anything realized OOS.

Optimized drawdown 31.7% breached the hypothesis's own pre-registered 25% cap; 3 of 6 pre-registered predictions missed (OOS Sharpe, drawdown, profit factor).

lookback is a cliff parameter (Sharpe 0.20-0.63 across +-20%) — the one param the sweep leaned on is unstable.

Benjamini-Hochberg at q=0.10 over 269 programme candidates keeps 4. A candidate that does not survive here is not distinguishable from the programme's own noise, however good its individual statistics look.

SolIntermediateSkipWindowTsmomLS-21193df43b

This strategy tried to fix the sub-fee whipsaw that killed prior SOL trend clones by measuring 4H momentum over an intermediate window that excludes the most recent reversal-prone bars, sizing inverse to volatility and exiting on an ATR trail. Its initial backtest passed the pre-optimization gate on a decisive 362-trade sample (Sharpe 0.64, PF 1.24, per-trade edge 1.17% well above trading costs), earning a full optimization run. But the 3-phase deflation was decisive: in-sample Sharpe 2.96 fell to OOS -0.099, the holdout printed -0.228 on 59 trades, PBO 0.51 and deflated Sharpe 0.0 marked the selected result as noise, and the optimized drawdown of 31.7% broke the 25% pre-registered cap. The analyst abandoned it after one iteration as an overfit single-name OHLCV TSMOM clone with no robust region to iterate toward.

Reframing single-name OHLCV trend-following (here as a skip-window intermediate momentum) adds no robustness that the 3-phase deflation cannot strip; a strong in-sample Sharpe on a slow single-instrument strategy is indistinguishable from best-of-N noise, and lookback being a cliff parameter means there is no robust region to iterate toward.

The analyst abandoned it after optimization on multiple HARD gate failures — is_overfitted=true, negative OOS/holdout Sharpe, PBO>0.5, DSR 0.0 / not significant / fails programme FDR — and the optimized 31.7% drawdown breached the hypothesis's own pre-registered 25% cap.

A long-short, single-instrument 4H time-series momentum strategy on SOLUSDT.BINANCE USD-M that measured trend over an intermediate skip-window (t-N..t-k, excluding the most recent k bars) to strip short-term reversal whipsaw, with inverse-volatility sizing and an ATR-trailing exit.

The initial backtest over 362 trades and a 2183-day window looked viable (Sharpe 0.64, PF 1.24, 22.2% max drawdown, avg_trade_return_pct 1.17% above the 0.15% fee floor), but 3-phase optimization exposed it: in-sample Sharpe 2.96 collapsed to out-of-sample -0.099, holdout Sharpe was -0.228 over 59 trades, PBO 0.51, and deflated Sharpe 0.0.