Strategies

BTC Intraday Impulse-Bar Continuation (Single-Instrument USD-M, Long-Short, 1H, Enter on High-Conviction Momentum Thrust, Trailing-Stop Continuation, Time-Stop, Discrete Capital-Capped, Pure OHLCV)

Outcome: Abandoned

BtcImpulseBarContinuationTrendLS1H

Outcome Summary

Built as the BTC arm of the analyst-endorsed intraday impulse-continuation family, this strategy copied the promoted ETH-1H sibling's template exactly, betting that BTC's depth and low mania-beta would give the cleanest impulse signal. In backtest it was only barely profitable — Sharpe 0.134 against the sibling's 1.40, profit factor 1.054, +$36/trade — and carried heavy cost drag with capacity near $528K. Worse, the edge had decayed forward: a strong 2024 gave way to -17.9% in 2025 and -11.8% in 2026 with rolling Sharpe running -4 to -6, so any parameter tuning would land its holdout on a losing regime. The analyst abandoned it at the pre-optimization gate after 2 iterations, noting the same signature had already sunk the family's 4H variant.

Strategy report

Hypothesis

A LONG-SHORT, SINGLE-INSTRUMENT, pure-OHLCV intraday MOMENTUM-CONTINUATION strategy on BTCUSDT.BINANCE USD-M perpetual futures (1-HOUR bars). It extends the factory's ONE analyst-endorsed open direction: the GENERALIZING intraday IMPULSE-CONTINUATION family (the in-pipeline EthImpulseBarContinuationTrendLS1H, cited repeatedly as generalizing). The analyst's standing instruction after four consecutive basket-refinement failures is explicit: 'STOP generating basket refinements/overlays/signal-swaps; redirect ALL novelty to genuinely different mechanisms, e.g. the generalizing intraday impulse-continuation family.' This is that family, applied to BTC. MECHANISM: an IMPULSE bar is a 1H bar whose directional thrust (close-to-close return, or body) exceeds impulse_mult x ATR(atr_period) AND closes in the strong third of its range -- a high-conviction momentum event (an information/flow thrust, NOT noise). The strategy enters in the impulse direction to capture the short-horizon CONTINUATION before the move is fully absorbed. WHY IT GENERALIZES WHERE OTHER SINGLE-ASSET INTRADAY DECAYED: VWAP-continuation decayed (2023-2026) because it rode EVERY excursion (noisy, continuous); impulse-continuation is EVENT-DRIVEN and SELECTIVE -- it only fires on rare, large, high-conviction bars, which is the higher-quality signal that makes the ETH sibling generalize. It is also NOT single-asset trend/TSM (dead: continuous trend-sign whipsaws and is mania-beta) -- it is discrete large-bar continuation. WHY BTC: the DEEPEST, most-liquid major -> the cleanest impulse signal (least idiosyncratic noise), the lowest mania-beta (the killer of every single-asset alt), and dense 1H events. WHAT IT AVOIDS (every closed path): NOT a basket refinement (the prohibited axis -- 4 consecutive degradations); NOT funding/carry (uncreditable in the engine); NOT reversion (dead); NOT a continuous-rebalance/netting-flip strategy (the 316%/385% sizing-artifact cause -- this uses DISCRETE enter-then-exit-to-FLAT entries, like the cleanly-sized ETH impulse sibling). EVERY design choice respects the lessons: (1) DISCRETE entries, flat between signals, sized against STABLE realized equity (excl. unrealized PnL), hard-capped <=100% (leverage 1x) -- avoids the inflation artifact. (2) FEE-AWARE: fires only on large-impulse bars, targeting continuation moves well above the ~0.10% RT futures floor -- NOT scalping. (3) DENSE: ~150-300 impulse events/yr at 1H -> populates the 3-window walk-forward (no event-sparsity). (4) PURE OHLCV: no ticks (2-day wall), options (unwired), or funding (data ends 2024). Per the 'favor simpler, fewer parameters' mandate, 3 tunables (impulse_mult, trail_mult, time_stop_bars).

Iteration 2 fixes ONLY the reported Layer-3 crash ('float' object is not callable, 0 bars processed). Root cause, confirmed against sibling strategies in src/strategies/generated (e.g. BtcHyperliquidImpulseBarContinuationTrendLS1H_v4): the strategy stored the trailing-stop price in self._stop, which shadows NautilusTrader's reserved Component._stop lifecycle method; when the engine invokes self._stop() the float attribute is called and the backtest aborts before any bar is processed. The fix is a pure mechanical rename of self._stop to self._stop_px at all six occurrences (__init__, calculate_signal trailing-stop updates, should_enter initial-stop set, should_exit stop checks). No signal logic, imports, sizing, thresholds, config, or parameters changed, so Layer 1 (static) and Layer 2 (synthetic) remain green and the impulse-continuation edge is untouched.

Pre-optimization gate: the BTC arm of the impulse-continuation family does not inherit the promoted ETH-1H sibling's edge, and the binding failures are structural (not tunable). The base Sharpe is 0.134 (vs ETH 1.40) with PF 1.054, expectancy +$36/trade, and sharpe_ci_low -0.86 deeply through zero — barely profitable, not a modest edge. Decisively, the edge has DECAYED forward: after 2024 (+23.0%), 2025 is -17.9% and 2026 -11.8%, with rolling Sharpe running -4 to -6 continuously through 2025-2026 — so the optimization holdout (most-recent window) will land on this decayed regime and fail regardless of parameter choice. On top of that, cost drag is heavy (impact_cost_pct 43.5%, commission 11.53% of gross, capacity only ~$528K), making the thin per-trade edge fee-fragile. This is the same signature that already got the 4H variant of this exact family (EthFourHourImpulseBarContinuationTrendLS4H) abandoned — a modest continuation edge whose untouched recent holdout goes negative. Mirroring a promoted sibling's code EXACTLY does not transfer promotability to a different asset whose signal is weaker and whose recent regime is losing. Optimizing 3 params on a 0.134-Sharpe, recently-negative, cost-heavy book is not worth 2 hours.

Implementation

Single-instrument long/short intraday impulse-bar CONTINUATION on BTCUSDT.BINANCE USD-M perpetual, 1-HOUR bars, pure OHLCV. Detects a large-impulse bar (current-bar return expressed as a z-score of the last vol_lookback bar returns) and enters WITH the impulse: LONG on impulse >= +impulse_z, SHORT on impulse <= -impulse_z. Exits via a tight initial ATR stop, an ATR trailing stop that ratchets to let winners run, or a time cap. Discrete enter-then-exit-to-flat, one position at a time, risk-based sizing capped at max_notional_pct*leverage of stable equity (leverage 1x).

Verification Results

Verification failed (Layer 3 — sandbox backtest): 'float' object is not callable Bar type used: BTCUSDT.BINANCE-1-HOUR-LAST-EXTERNAL, Bars processed: 0 Ensure your strategy produces trades with the given data and parameters.

Backtest Review

Sharpe: 0.13
Total return: 34.88%
Max drawdown: 28.35%
Trades: 855
Win rate: 36.4%
Profit factor: 1.05

Mechanism implements the hypothesis: 433 long / 422 short discrete enter-to-flat trades, impulse-z entries with trailing stops, dense sample (855 trades), favorable win/loss asymmetry (~1.8:1)

Genuinely low-net-exposure single-instrument book (beta 0.02, exposure 21%), no sizing artifact

Part of an analyst-endorsed family with a promoted ETH-1H sibling

Very weak base: Sharpe 0.134 (vs ETH sibling 1.40), Sortino 0.293, PF 1.054, expectancy +$36/trade, sharpe_ci_low -0.86 straddling 0, CAGR 4.27%

Decisive recent decay: 2024 +23% then 2025 -17.9% and 2026 -11.8%, with rolling Sharpe -4 to -6 throughout 2025-2026 — the recent holdout window will fail (forward-viability problem, not tunable)

Heavy cost drag: impact_cost_pct 43.5%, commission 11.53% of gross, capacity only ~$528K — thin edge is fee/impact-fragile

Same signature as the already-abandoned 4H variant of this family (modest continuation edge, negative recent holdout as 2025-2026 decays) — asset/cadence-selectivity failure, not a promotable transfer

Outcome Summary

Mirroring a promoted sibling's exact code does not transfer promotability to a different asset — BTC's impulse-continuation signal is weaker than ETH's, its recent regime is losing, and a thin per-trade edge is too fee- and impact-fragile to survive.

The analyst issued an 'abandon' verdict at the pre-optimization backtest-review gate: the edge is structurally weak and has decayed forward (2024 +23%, then 2025 -17.9% and 2026 -11.8% with rolling Sharpe -4 to -6), so the most-recent optimization holdout window would fail regardless of parameters. Optimization and all later stages were never reached.

A single-instrument, long-short intraday momentum-continuation strategy on BTCUSDT.BINANCE USD-M 1H futures that fires only on rare high-conviction impulse bars (return z-score beyond a threshold) and rides the continuation with a trailing stop, discrete enter-to-flat, mirroring the promoted ETH-1H impulse-continuation sibling on a deeper, lower-mania-beta major.

Over ~5.7 years (2019-12-31 to 2026-06-25) it took 855 discrete trades (433 long / 422 short) with a low 36.4% win rate but favorable ~1.8:1 win/loss asymmetry, ending at +34.9% total return, CAGR 4.27%, profit factor 1.054, expectancy +$36/trade, and a 28.3% max drawdown. The Sharpe was only 0.134 (CI -0.86 to 1.01) versus the ETH sibling's 1.40, with heavy cost drag (impact 43.5%, commission 11.53% of gross, capacity ~$528K).

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