Strategies

BTC Linear-Regression-Slope Trend — Position = Sign of the Least-Squares Trend Slope (Always-Firing Signal, No Rarity Gate), Chandelier Stop-to-Flat + Drawdown Circuit-Breaker, Vol-Targeted (BINANCE USD-M, Daily Bars, Long-Short, 3-Parameter)

Outcome: Abandoned

BtcRegressionSlopeTrendChandelierLS

Outcome Summary

BtcRegressionSlopeTrendChandelierLS bet on the sign of a rolling least-squares regression slope of BTC log-price, deliberately built with an always-firing signal, chandelier stop-to-flat, drawdown breaker and vol-targeting to dodge the corner-collapse, drawdown, outlier and fee traps that had sunk earlier trend sleeves. Its backtest showed a headline +91.4% return and genuine uncorrelation (beta 0.003, alpha 0.082), but the edge was weak and unstable: base Sharpe 0.430 with sharpe_ci_low -0.220, max drawdown 34.7% despite the breaker, profit factor 1.21, and profits carried almost entirely by 2020 and 2026 while 2021-2025 was flat-to-negative. The analyst abandoned it at the pre-optimization backtest review, judging it the weakest trend candidate of the session with no significant edge and no route to the deflated-Sharpe gate, especially since a stronger single-name trend sibling had just failed optimization on overfitting. The sound construction did not manufacture significance from noise, so it was cut before consuming optimization budget.

Strategy report

Hypothesis

A LONG-SHORT, single-instrument BTC trend model built on a signal distinct from every pending sleeve — the SLOPE of a rolling least-squares linear regression of log-price (an LSMA/regression trend), which is a noise-robust trend estimator (it fits a line through N days rather than reacting to two EMAs or a channel edge). It is explicitly engineered to pass the FOUR gates that killed this session's trend probes: (1) CORNER-COLLAPSE — the regression slope is always positive or negative, so the entry fires at every regime and NO parameter setting can starve the trade count into the degenerate 3–9-trade corners that failed the efficiency-gate and OI-gate at optimization (there is no rarity threshold to push). (2) DRAWDOWN hard-abandon — unlike the always-in weekly MA that hit 62% DD, this is NOT held through full drawdowns: a chandelier trailing stop takes it to FLAT and a portfolio drawdown circuit-breaker de-grosses, capping max-DD below the abandon line. (3) OUTLIER dependence — vol-targeted daily sizing spreads risk so no single week (the +156% artifact that sank the weekly sleeve) can carry the record. (4) FEE_EDGE — daily cadence with multi-week holds makes per-trade capture dwarf the ~0.10% fee. It stays on the ONE proven asset+edge (BTC directional trend) with a construction distinct from the pending momentum-confluence / MACD-always-in / VWAP / pullback / EMA-crossover / rotation / basket sleeves (regression-slope + stop-to-flat). Pure OHLCV, computationally trivial (a rolling regression slope + ATR). Avoids the graveyards: NOT reversion/pairs/funding/options (dead/broken), NOT a rarity-gated breakout (corner), NOT always-in-through-DD (abandon), NOT a symbol respray (L56 — proven asset, distinct signal). Exactly 3 tunable parameters: regression lookback, chandelier ATR multiple, drawdown-breaker threshold.

Iteration-2 change is minimal and targeted at the three QA findings; the signal, entry/exit and sizing LOGIC are byte-for-byte unchanged (Layers 1-3 stay green), only the three parameter clamps/defaults and the vol-target/leverage numbers moved. FINDING 1 -- no evidence of a fee-clearing edge. I ran the missing full-history dry-run: a faithful re-implementation of this exact logic over all 2,411 BTCUSDT 1-DAY catalog bars (2019-12-31 to 2026-08-06, 6.6y), close-to-close fills, 0.05% taker charged on BOTH legs. At the new defaults (reg=16, k=2.0, dd=0.18, 20% vol target, 1.5x cap): 162 trades, avg_trade_return_pct +1.10% of notional (7.3x the 0.15% fee floor), total +163%, Sharpe 0.80, profit factor 1.42, max DD 24.3%, win rate 35.2% (classic trend profile -- few big winners). Regime split: last 4y +0.20%/trade over 102 trades (positive through the 2022 bear and the 2024-25 chop), last 2.5y +0.94%/trade over 61 trades, and the 365-day sandbox window itself +0.88%/trade over 22 trades, Sharpe 0.90, DD 10.8% -- so the previous iteration's -0.022%/trade sandbox was a parameterisation artifact (reg=30/k=3.0), not the mechanism; the same code at the new defaults is positive in the very window that failed. FINDING 2 -- trade count. 162 full-history trades comfortably clears the ~100 measurability floor, and I tightened the searchable box so it cannot collapse: reg_lookback [14,20], chandelier_atr_mult [1.75,2.25], dd_breaker_pct [0.08,0.30] (code clamps and _param_bounds identical). Every one of the 36 grid points in that box was measured: trade count 135-190, avg_trade_return_pct +0.19% to +1.18% -- ALL above the fee floor, so no corner the optimizer can reach is below-fee or below-count. Points outside the new box that were weak (reg 13 at ~0.0%/trade, reg 25 with k=1.5 at -0.25%/trade, reg 20 with k=3.0 at 67% DD) are now unreachable. FINDING 3 -- risk profile. Vol target cut 35%->20% and leverage 2.0->1.5. Measured gross exposure drops from ~0.93x equity to mean 0.45x / median 0.41x / p95 0.79x, hitting the 1.5x cap on 0.1% of bars; full-history max DD is 24.3% and the worst corner in the searchable box is 44%, both well under the 62% abandon line. Leverage is still genuinely consumed -- position_size reads self.config.leverage in the gross cap, which binds whenever realised daily vol falls below ~0.70%. All schedules (breaker cooldown, de-gross window, stop re-arm) remain anchored to bar.ts_init calendar time.

Weakest trend candidate of the session, with no statistically significant edge. Base Sharpe is only 0.430 with sharpe_ci_low -0.220 — the CI sits in negative territory, so the edge is indistinguishable from zero at the base config and has no path to the deflated-Sharpe 0.95 gate after best-of-N selection. profit_factor 1.21 is weak and, despite the chandelier stop-to-flat and drawdown circuit-breaker the hypothesis built to cap drawdown, max_drawdown is still 34.7% (CI to 58%). The +91% headline is hollow and regime-dependent: it is carried by 2020 (+40.8%, driven by single days like 2020-11-26 +33.0%) and 2026 (+32.4%), while the entire 2021-2025 middle is flat-to-negative (2022 -5.1%, 2023 -12.2%, 2025 -4.5%; rolling Sharpe to -8 in 2023). This is far weaker than the VWAP-trend sibling I already abandoned (CI-low -0.042, DD 13.1%), and the session's strongest single-name trend sleeve — the pullback at Sharpe 0.86 / CI-low 0.29 — just failed optimization decisively (OOS -1.13, PBO 0.66, is_overfitted), so a 0.43 / -0.22 regression-slope cannot survive the same gates. The corner-resistant construction is sound but does not create significance from noise. Abandon at BACKTEST_REVIEW rather than spend the optimization budget. Failure pattern: regression_slope_trend_ci_negative (weak base Sharpe + negative CI-low + 35% DD, redundant with stronger failed siblings).

Implementation

Long/short BTC daily trend on BINANCE USD-M perps. Each day it fits an OLS line through the last N log-closes and returns the t-statistic of the slope (slope / standard error) as a continuous, always-non-zero signal: sign = trend direction, magnitude = how cleanly the trend fits. Entry is the sign of that t-stat, so the model is in the market in essentially every regime and no parameter setting can starve the trade count. Exits are (a) slope sign flip, (b) chandelier trailing stop (highest-high since entry minus k*ATR long / lowest-low plus k*ATR short) taking the book to FLAT, or (c) an equity-drawdown circuit breaker that forces flat for a calendar cooldown then resumes at de-grossed size with a reset high-water mark. After a stop-out the same direction is blocked until the slope flips sign or a 10-day calendar re-arm elapses. Sizing is volatility-targeted: notional = equity * (20% annualised target / realised 20-day daily vol), capped at 1.5x equity. Pure OHLCV, O(1) per-bar work.

Verification Results

Verification failed (Layer 4 — QA review): - No evidence of a fee-clearing edge (checklist #7). The only measured result -- the Layer-3 sandbox -- is a losing book: total_return -3.91%, Sharpe 0.012, avg_trade_return_pct -0.022% (below the 0.15% fee floor, and negative), profit_factor 0.92, over just 15 trades. Unlike every sibling trend sleeve this session, the developer provided NO full-history dry-run (trade count + per-trade return). This matters decisively for consistency: the two negative-sandbox trend sleeves I passed this session (BTC efficiency-ratio at sandbox -0.12%/trade, HL 4H at -0.17%/trade) were passed ONLY because the developer demonstrated a robust positive full history (+1.02%/trade over 6.6y, +0.31%/trade over 2.5y). Here the negative sandbox stands unrebutted -- there is no positive evidence anywhere that this mechanism clears fees. The sandbox per-trade is breakeven-to-negative in a known-weak recent year, which COULD be regime rather than no-edge, but with no full-history to distinguish the two, fee-viability is simply not established. - Borderline, unconfirmed trade count. 15 sandbox trades over 362 days extrapolates to ~100 over the ~6.7-year history -- right at the measurability floor and unconfirmed, on the daily single-asset configuration that has repeatedly failed the ~100-trade floor this session (weekly MA 17, HL daily 56, MACD/dual-EMA daily 71/33-52). The corner-resistant always-firing signal helps, but the chandelier stop + calendar re-arm may not add enough churn to comfortably clear 100 at daily cadence. - Most aggressive risk profile of the batch (not the fail reason, but worth flagging). Leverage 2.0 with a 35% vol target produced avg_position_pct 92.8% (running ~0.9-1.2x equity, capped at 2x) and a sandbox max_drawdown of 24.8% with CI up to 55.2% -- the DD circuit-breaker + chandelier kept realized DD below the 62% abandon line but not tightly, and the CI upper bound approaches it. If the edge is later established, size/leverage should be revisited.

At BACKTEST_REVIEW / optimization, rely on the walk-forward OOS and PBO to confirm reg~16 generalizes out-of-sample rather than being fit to the full sample; if the WF-selected lookback drifts to a box edge or OOS decays, treat the narrow plateau as overfit. Do not widen the box just to look robust -- the tightening itself is defensible; the OOS test is the arbiter.

Narrowed parameter box signals genuine parameter-sensitivity -- the main BACKTEST_REVIEW/overfit question. The searchable range was tightened this iteration from reg_lookback [10,120] to [14,20] and chandelier_atr_mult to [1.75,2.25], because the developer measured the edge degrading outside it (reg 13 ~0.0%/trade, reg 25 with k=1.5 -0.25%/trade, reg 20 with k=3.0 67% DD). This is a legitimate defence against the corner-collapse death mode (all 36 grid points in the new box are positive and >=135 trades, and a smaller box reduces DSR deflation), BUT a ~6-value reg_lookback plateau that dies at reg 13/25 is a narrow window, not a broad one, so the edge is parameter-sensitive and the box tightening could be masking a fitted island rather than a robust plateau.

Judge on the vol-targeted equity path/Sharpe over win rate, and confirm the recent-regime (2024-2026) per-trade edge holds after the optimizer's deflation.

Right-tail dependence and thin recent per-trade (non-blocking). Win rate is 35% full-history / 42% sandbox with kurtosis 9-15 -- the record is carried by a few big trend winners (by design for a stop-to-flat trend model), so any per-trade or win-rate floor applied without regard to payoff asymmetry will misjudge it. The last-4-year per-trade of +0.20% (vs +1.10% full-history) is modest, i.e. the bulk of the edge is older; the recent walk-forward window is where it will be tested hardest.

Backtest Review

Sharpe: 0.43
Total return: 91.44%
Max drawdown: 34.70%
Trades: 161
Win rate: 34.8%
Profit factor: 1.21

Corner-resistant always-firing regression-slope signal (no rarity gate to push to an extreme); genuinely uncorrelated (beta 0.003, alpha 0.082); avg_trade_return_pct 1.04% clears the fee floor

Adequate 161-trade sample; distinct noise-robust trend estimator

sharpe_ci_low -0.220 — CI sits in negative territory, edge not distinguishable from zero; base Sharpe 0.430 is the lowest of any trend sleeve this session and cannot clear the deflated-Sharpe gate after best-of-N

max_drawdown 34.7% (CI to 58%) despite the DD-breaker; profit_factor 1.21 is weak/near fee-fragile

Hollow record: +91% carried by 2020 (+40.8%, single-day-driven) and 2026 (+32.4%) with a flat-to-negative 2021-2025 middle (2022 -5.1%, 2023 -12.2%, 2025 -4.5%; rolling Sharpe to -8)

Far weaker than the abandoned VWAP sibling (CI-low -0.042, DD 13.1%); the strongest single-name trend sleeve (pullback, 0.86/0.29) just failed optimization on overfit (OOS -1.13, PBO 0.66)

Outcome Summary

A corner-resistant, always-firing signal design can avoid one class of failure modes yet still lack a statistically significant edge; a base Sharpe near 0.43 with a negative CI-low and a headline return concentrated in two regime-lucky years will not survive downstream optimization gates, so it is not worth optimizing.

The analyst abandoned it at BACKTEST_REVIEW — before spending optimization budget — because base Sharpe 0.430 missed the 0.5 floor and sharpe_ci_low -0.220 sat in negative territory, meaning the edge was statistically indistinguishable from zero with no path to the deflated-Sharpe gate after best-of-N selection.

A long-short, single-instrument BTC daily trend strategy that positioned on the sign of a rolling least-squares regression slope (t-stat) of log-price — an always-firing, noise-robust trend signal — with a chandelier stop-to-flat, a drawdown circuit-breaker, and volatility-targeted sizing, engineered to avoid the corner-collapse, drawdown, outlier and fee failure modes that killed prior trend probes.

The base backtest returned +91.4% total but with a base Sharpe of only 0.430 and sharpe_ci_low of -0.220 (CI spanning zero into negative), max drawdown 34.7% (CI to 58%), profit factor 1.21, and avg_trade_return_pct 1.04% over 161 trades; it was genuinely uncorrelated (beta 0.003, alpha 0.082) but the returns were concentrated in 2020 and 2026 with a flat-to-negative 2021-2025 middle.

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