DOT Dual-Timeframe Momentum – Backtest

Hypothesis

A LONG-SHORT, single-instrument, pure-OHLCV momentum-confluence trend follower on DOTUSDT.BINANCE (USD-M perpetual): 4H primary momentum, 1D trend confirm. This is a DELIBERATE STRUCTURAL TWIN of the factory's ONLY empirically-surviving mechanism family — the BTC and ADA Dual-Timeframe Momentum Confluence strategies now in paper_stage (Sharpe 1.99 and 1.60) and the ETH Volume-Confirmed variant (Sharpe 1.05). Those survivors share one recipe: take a directional position ONLY when a fast (4H) trend and a slow (1D)……Show moreShow less

A LONG-SHORT, single-instrument, pure-OHLCV momentum-confluence trend follower on DOTUSDT.BINANCE (USD-M perpetual): 4H primary momentum, 1D trend confirm. This is a DELIBERATE STRUCTURAL TWIN of the factory's ONLY empirically-surviving mechanism family — the BTC and ADA Dual-Timeframe Momentum Confluence strategies now in paper_stage (Sharpe 1.99 and 1.60) and the ETH Volume-Confirmed variant (Sharpe 1.05). Those survivors share one recipe: take a directional position ONLY when a fast (4H) trend and a slow (1D) trend AGREE, stand flat on disagreement, trail winners with an ATR stop, and expose only ~3 parameters. The recent DEATHS of this family (BNB 8H+1D, AVAX 4H+volume+close-location filter, SOL-HL) all died overfit/PBO because they ADDED filters/timeframes/parameters that curve-fit the search — NOT because the core mechanism is fee-dead (every one cleared the per-trade fee floor comfortably: 1.36%, 0.53%, etc.). The lesson from the survivor-vs-death split is UNAMBIGUOUS: keep the recipe minimal and move it to an UNDER-MINED liquid instrument, do not add cleverness. This hypothesis does exactly that: identical minimal structure, applied to DOT — a liquid ~6-year-history USD-M perp that is essentially absent from the corpus (62% of recent research is BTC; DOT does not appear in any recent failure list), and expressed LONG-SHORT to fill the under-represented direction bucket (long_short is only 14.2% of the portfolio vs an 85.8% long-only glut).

The hypothesis asks for a deliberate structural twin of the surviving BTC/ADA Dual-Timeframe Momentum Confluence strategies, moved to an under-mined liquid instrument (DOT) and expressed long-short. I copied the ADA survivor's exact structure rather than inventing anything: 4H primary + genuine 1-DAY extra bar leg (not an EMA proxy on 4H bars, since the hypothesis names a 1D trend confirm), flat-on-disagreement, ATR trailing exit, three tunables with declared bounds. No volume filter, no close-location filter, no……Show moreShow less

The hypothesis asks for a deliberate structural twin of the surviving BTC/ADA Dual-Timeframe Momentum Confluence strategies, moved to an under-mined liquid instrument (DOT) and expressed long-short. I copied the ADA survivor's exact structure rather than inventing anything: 4H primary + genuine 1-DAY extra bar leg (not an EMA proxy on 4H bars, since the hypothesis names a 1D trend confirm), flat-on-disagreement, ATR trailing exit, three tunables with declared bounds. No volume filter, no close-location filter, no third timeframe, no regime gate -- those additions are exactly what the hypothesis identifies as the cause of the family's overfit deaths, so I added none. Venue is BINANCE USD-M because the strategy shorts and uses leverage>1; leverage 2.0 is consumed in position_size via the max_notional cap so it is not inert. extra_instruments carries DOTUSDT.BINANCE positionally paired with the daily bar type so the slow leg actually loads; the daily leg is a slower timeframe so it does not trigger the same-timeframe sync barrier, and its lookback (18 daily bars) plus the 4H warmup (~77 bars, ~13 days) is a small fraction of any sandbox window. Entry is a two-condition agreement (not a 3+ condition conjunction), which fires often on a trending/chopping alt and should produce hundreds of trades over DOT's ~6 years of history; per-trade capture is trend-length scaled by a 3-ATR trail, comfortably above the ~0.10% round-trip taker cost.

Overfit / no-generalization death (NOT fee-edge — per-trade return 0.69% clears the floor). Three unwaivable HARD gates fail: walk-forward is_overfitted=TRUE (avg IS Sharpe 1.339 -> avg OOS 0.099, windows +0.14/-0.85/+1.01), PBO 0.541 > 0.5, and the holdout FAILED with Sharpe -0.266 <= 0 on 50 trades over an unseen 331-day window. Robustness confirms the selected Sharpe is best-of-N noise: deflated_sharpe 0.119 << 0.95, is_significant=false, Sharpe CI [-0.42, 1.18] straddles zero, and it does not survive……Show moreShow less

Overfit / no-generalization death (NOT fee-edge — per-trade return 0.69% clears the floor). Three unwaivable HARD gates fail: walk-forward is_overfitted=TRUE (avg IS Sharpe 1.339 -> avg OOS 0.099, windows +0.14/-0.85/+1.01), PBO 0.541 > 0.5, and the holdout FAILED with Sharpe -0.266 <= 0 on 50 trades over an unseen 331-day window. Robustness confirms the selected Sharpe is best-of-N noise: deflated_sharpe 0.119 << 0.95, is_significant=false, Sharpe CI [-0.42, 1.18] straddles zero, and it does not survive programme-level FDR (keeps 4 of 242, candidate_p 0.88). Per the overfit-vs-dead-premise rule, high-IS/OOS-collapse is ABANDON not revise: this exact DOT dual-TF confluence has ALREADY died once (abandon_code fee_edge, similarity 0.812), and every sibling that pushed the minimal recipe onto a new horizon/instrument died overfit — BTC faster-horizon (overfit), BTC macro-confirmed (overfit, 3 iters), BTC triple-TF (overfit, 2 iters), DOT daily TSMOM (61% DD). Only the original BTC/ADA 4H+1D configurations survived; porting to DOT does not add robustness the deflation can strip, it just overfits again. There is no un-explored robust parameter region to tune toward (sensitivity shows 3 cliff params — fast_ema, ema_ratio, daily_lookback — and the whole grid tops out at ~0.34 Sharpe), and the oos_vault is itself flagged UNDERPOWERED (~73 trades vs 100 needed), so the mechanism is too slow to validate on DOT's history. Failure pattern: overfit / no-generalization (single-name dual-TF momentum-confluence clone).

Implementation

Long-short dual-timeframe momentum confluence on DOTUSDT.BINANCE USD-M perpetual. The fast leg (4H primary bars) computes an ATR-normalized EMA spread (EMA12 vs EMA~36) returned continuously every bar; the slow leg (1-DAY bars via extra_instruments/extra_bar_types) computes an 18-day return sign as the trend confirm. A position is taken ONLY when the two agree (spread>0 and daily up -> LONG; spread<0 and daily down -> SHORT); on disagreement the strategy stands flat. Exits on (a) an ATR trailing stop of 3x the entry ATR from the best close since entry, (b) the daily trend flipping against the position, or (c) the 4H spread flipping against the position. Sizing is risk-first: equity * 1.5% / (trail_atr_mult * ATR), hard-capped at 25% of current equity times the requested 2x leverage, so per-trade risk and gross notional are both bounded relative to current equity. Only three tunables are exposed (ema_ratio, daily_lookback, trail_atr_mult), each clamped to the ranges declared in _param_bounds.

Verification Results

No change required; the overrides are intentional and call through to the base where needed.

min_bars_required() and _submit_exit() override framework methods (base_shadow warnings from static analysis). Both overrides are deliberate and behave correctly: min_bars_required gates warmup and _submit_exit calls super() then resets trailing/side state. No functional problem.

Sandbox (unoptimized smoke test) is only mildly positive: Sharpe 0.33, and the normal vol tercile is net negative (-7.1%, Sharpe -0.15) with the entire edge concentrated in the stressed tercile (+37.6%, Sharpe 1.08). That regime concentration is expected of a trend follower, but the analyst should confirm on full history that the DOT carrier reproduces the BTC/ADA survivor's Sharpe rather than relying on a single high-vol regime. Per-trade return 0.85% and 390 trades are healthy, so measurability is not in doubt.

Backtest Review

Sharpe: 0.33
Total return: 39.72%
Max drawdown: 15.66%
Trades: 390
Win rate: 36.9%
Profit factor: 1.09

Mechanism correctly implements the hypothesis: 191 long / 199 short is the intended balanced long-short confluence, flat-on-disagreement, ATR-trailing exit — this IS the factory's only empirically-surviving recipe (BTC/ADA dual-TF confluence) applied minimally to an under-mined instrument, no added filters.

Numerically viable on a reliable 390-trade / 6-year sample: avg_trade_return_pct 0.849% sits far above the 0.15% USD-M fee floor (commission only 3.92% of gross), so this is NOT a fee-edge case; expectancy +$61/trade, PF 1.09 > 1, drawdown a modest 15.66%.

DOT is essentially absent from the corpus and long_short is an under-represented direction bucket, so this is a genuine breadth addition rather than a re-test of a dead target.

Thin base edge: Sharpe only 0.335 with the bootstrap CI [-0.51, 1.09] straddling zero, and PF 1.09 sits in the fee-drag band. This is well below the ADA/BTC survivor tier (Sharpe ~1.6-2.0, PF ~2.2) — the real risk is that the 3-phase deflation strips it to noise.

Regime-concentrated: the entire edge lives in the stressed vol tercile (+37.6%, Sharpe 1.08) while the normal tercile is net negative (-7.1%); recent calendar years are negative (2025 -1.8%, 2026 -9.6%). QA's concern is that the carrier is riding one high-vol regime rather than reproducing the survivor's broad-based edge — optimization must confirm a robust parameter region, not just the current one.

Analysis

Per-trade edge clears the fee floor comfortably (avg_trade_return_pct 0.69%, well above the 0.15% Binance USD-M floor) so this is NOT a fee-edge death.

Mechanism is correctly implemented: balanced 162 long / 184 short confluence, flat-on-disagreement, ATR-trailing exit — a faithful minimal twin of the surviving BTC/ADA dual-TF recipe.

Concentrates positive performance in the stressed vol tercile (Sharpe 1.49), the regime a trend follower should exploit.

Walk-forward is_overfitted=TRUE: avg IS Sharpe 1.339 collapses to avg OOS 0.099, with OOS windows +0.14 / -0.85 / +1.01 — no consistent out-of-sample edge.

PBO 0.541 > 0.5 — the parameter selection is more likely than not overfitting.

Holdout FAILED (HARD gate): holdout Sharpe -0.266 <= 0 on 50 trades over an unseen 331-day window — the edge did not show up on the last 20% of history.

Robustness collapses: deflated_sharpe 0.119 << 0.95, is_significant=false, Sharpe CI [-0.42, 1.18] straddles zero, and it does not survive programme-level FDR (keeps 4 of 242, candidate_p 0.88).

Sensitivity shows 3 cliff parameters (fast_ema, ema_ratio, daily_lookback) and the whole grid maxes at ~0.34 Sharpe — no robust parameter region to tune toward.

The 'normal' vol tercile is persistently negative (-19.3%, Sharpe -0.65), so the headline depends on a single regime.

oos_vault flagged UNDERPOWERED (~73 expected trades vs 100 needed) — the mechanism is too slow to validate on this instrument's history.

Benjamini-Hochberg at q=0.10 over 242 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.

Outcome Summary

Three failed HARD gates post-optimization: walk-forward overfit (IS 1.34 → OOS 0.099), PBO 0.541 > 0.5, failed holdout (Sharpe -0.266).

A minimal 3-parameter long-short dual-timeframe momentum-confluence trend follower on DOTUSDT.BINANCE USD-M...

6-year, 390-trade backtest cleared the fee floor (0.85% per-trade, PF 1.09, 15.66% DD) but thin Sharpe 0.335; optimized to Sharpe 0.41 yet stayed overfit.