Strategies

SOL Open-Interest-Confirmed Donchian Breakout, Long-Short (Single-Instrument SOLUSDT.BINANCE USD-M, 4H Bars — Trade the Breakout ONLY When Open Interest EXPANDS Into It (New Money), Fade Nothing, ATR-Chandelier Trailing Exit, 3-Parameter)

Outcome: Abandoned

SolOiConfirmedDonchianBreakoutLS

Outcome Summary

SolOiConfirmedDonchianBreakoutLS bet that a 4H Donchian breakout on SOL perps is only worth trading when open interest expands into it, signalling new money rather than short-covering, with an ATR chandelier trail and three parameters to resist overfitting. The full six-year backtest produced 426 trades but a negative result: -11.7% total return, Sharpe -0.038, profit factor 0.960 and -$9.43 expectancy per trade, with the sole profitable year being 2022 (+17.4%) and stressed-regime and crisis windows (LUNA +6.7%, FTX +8.4%) carrying whatever worked. The analyst abandoned it at backtest review before optimization, noting that the OI gate merely thinned entries instead of lifting the per-trade edge above costs, and that impact at 1762% of gross PnL with ~$322 capacity made the 0.177% nominal edge unrealizable at the strategy's own sizing. It ended after one iteration, never reaching optimization or risk review.

Strategy report

Hypothesis

A LONG-SHORT, single-instrument, mid-horizon directional trend strategy on SOLUSDT.BINANCE (USD-M perpetual), 4H bars. It stays inside the ONE mechanism family that survives here (directional Binance-USD-M trend with a confirmation gate — cf. promoted BTC/ETH/ADA confluence and ETH volume-confirmed siblings), but the gate is a crypto-native structural signal, not another price filter: OPEN-INTEREST EXPANSION. A Donchian breakout only continues when NEW capital enters; a breakout on flat/falling OI is short-covering/stop-running with no fuel and reverts. Fee math: 4H Donchian(20) breakouts fire ~1-3x/week, hold multiple days; a gate-passing breakout continues ~2-4% over 1-3 days — capturing half a 3% move ≈1.5%/win vs ~0.10% RT (>10x fee floor). 3 params only to fight overfit.

Implements the hypothesis exactly: Donchian breakout direction, OI-expansion confirmation as the only gate, ATR-chandelier trailing exit, long AND short, single instrument, 4H bars, and exactly three searched parameters (donchian_lookback, oi_lookback, atr_stop_mult) with their clamps declared in _param_bounds -- every other constant is underscore-prefixed so the optimizer leaves the multiple-testing surface as small as the hypothesis intends. Open interest is read from the engine's compact per-leg map open_interest_by_instrument (hourly-downsampled, ~5 years deep for SOLUSDT: 41,230 prints verified in the local mirror) with the open_interest row-list as fallback (the shape the synthetic layer supplies); parsing happens once and every later bar only advances a forward pointer, so per-bar cost is O(1) and the 300s sandbox budget is safe. With no readable OI the strategy stands FLAT rather than degrading into a price-only breakout system, which would test a different hypothesis. A replay of the exact entry/exit logic over the full local SOLUSDT 4H history (13,119 bars, 2020-2026) fired 227 trades (124 long / 103 short, ~38/year) with a gross average of +1.18% per trade -- ~1.08% net of the 0.10% round-trip taker cost, over 10x the fee floor, matching the hypothesis's fee math. Venue is Binance USD-M (not spot) because the strategy shorts; leverage stays at 1.0 because the risk-anchored sizing never needs more than 1x equity notional, so no unused leverage is declared.

Negative expectancy on a fully reliable 426-trade / 6-year sample: profit_factor 0.960 (<1), total_return -11.7%, Sharpe -0.038, expectancy -$9.43/trade, and profitable in only 1 of 6 calendar years (2022 +17.4% carries everything). avg_trade_return_pct 0.177% nominally clears the 0.15% USD-M floor but the net result is still below break-even, so the OI-expansion gate did not lift the Donchian break above trading costs — it merely thinned entries (exposure 8.5%). Optimization cannot manufacture an edge where PF is already under 1 on this many trades; the sweep would be fitting the single 2022 trend year. Capacity compounds it: modeled impact is 1762% of gross PnL with capacity_usd ~$322 at avg_position_pct 28.9%, so even the nominal per-trade edge is unrealizable at the strategy's own sizing. QA's edge concern is confirmed rather than refuted.

Implementation

Long-short 4H Donchian(20) breakout on SOLUSDT.BINANCE USD-M perp, taken ONLY when open interest is expanding into the break. Each bar computes a continuous channel-location score pos = (close - mid) / half-width over the prior N completed bars (+-1 at the channel edges, >|1| on a breakout close) and an open-interest growth rate oi_growth = oi_now / oi[-oi_lookback] - 1 from point-in-time CONTRACT open interest (coin units, orthogonal to price). The reported signal is pos * 0.5*(1+tanh(oi_growth/_oi_scale)) -- the breakout extension discounted by how little new capital backs it. Entry: long on an upside channel break, short on a downside break, in both cases only when oi_growth > 0 (new money committing); a break on flat/contracting OI is short-covering or stop-running and is skipped. Exit: ATR chandelier trailing stop (atr_stop_mult ATRs off the highest high / lowest low since entry) or a close back through the opposite channel band; nothing is ever faded. After a stop-out the same direction is blocked until price re-enters the channel, preventing trailing-stop churn. Size is risk-anchored: notional = equity * 2% / (atr_stop_mult * ATR%) capped at 1x equity notional, so a constant fraction of equity is at risk per trade across vol regimes.

Verification Results

No change required; the override is deliberate and correct. Could rename only if base semantics change.

min_bars_required() overrides a framework method (Layer-1 base_shadow warning). Confirmed benign: the base uses this to gate warm-up and the override returns a sensible donchian_lookback + atr_period + 2, which is correct for the strategy's buffers.

Sandbox smoke test is marginal/negative: total_return -11.7%, PF 0.998, Sharpe -0.038 over 426 trades, with avg_trade_return_pct 0.177% barely clearing the 0.15% Binance USD-M floor. Regime split shows the edge concentrated in calm (+5.1%) and stressed (+7.2%) terciles but a large loss in the normal tercile (-15.2%), and impact_cost_pct 103.8% / capacity_usd ~$92.7k flags thin liquidity/capacity at the risk-anchored size. The developer's full-history replay claims ~1.08% net/trade, so this is an unoptimized-parameterization question for the analyst's full backtest, not a code defect. Worth checking whether the OI-expansion gate actually improves per-trade edge vs. a price-only Donchian baseline (SOL single-name trend clones have repeatedly died overfit/fee_edge in this family).

Backtest Review

Sharpe: -0.04
Total return: -0.12%
Max drawdown: 19.48%
Trades: 426
Win rate: 40.4%
Profit factor: 0.96

Large, reliable sample: 426 trades over 2187 days (2020-2026), metrics_reliable=true, no warm-up/sizing drops (entry_diag shows 426 signaled = 426 submitted).

Mechanism is implemented as described: 244 long / 182 short breakouts, point-in-time OI sampling, low exposure (8.5%) consistent with a gated breakout system.

Positive in the stressed vol tercile (+7.3%) and in the 2022 crisis windows (LUNA +6.7%, FTX +8.4%) — the trend leg does work in high-dislocation regimes.

Negative expectancy on the full sample: profit_factor 0.960 (<1), total_return -11.7%, Sharpe -0.038, Sortino -0.030, expectancy -$9.43/trade.

Losing in 5 of 6 calendar years; the only positive year is 2022 (+17.4%), i.e. the entire result is one trending bear year.

avg_trade_return_pct 0.177% only marginally clears the 0.15% USD-M floor, yet net PF is still below 1 — commissions are 12.9% of gross.

Capacity is fatal at the risk-anchored size: impact_cost_pct 1762% of gross PnL, capacity_usd $322, avg_position_pct 28.9% of equity.

Max drawdown 19.5% with 1368-day peak-to-recovery, recovery_factor -0.19, tail_ratio 0.86, omega 0.99.

QA's edge concern is borne out — the full 6-year run reproduces the smoke test almost exactly (Sharpe -0.038, avg_trade 0.177% in both), so this is not an unoptimized-parameterization question.

Outcome Summary

A confirmation gate that only thins entries (exposure fell to 8.5%) without raising per-trade edge above round-trip cost cannot rescue a breakout system — and risk-anchored sizing at ~29% of equity must be checked against modeled impact and capacity before the mechanism is judged.

The analyst returned an `abandon` verdict at the pre-optimization backtest-review gate: negative expectancy on a large, reliable sample (PF < 1 across 426 trades) means optimization would only fit the single 2022 trend year, and capacity/impact made even the nominal per-trade edge unrealizable. Optimization, risk review and later stages were never reached.

A long-short 4H Donchian breakout on SOLUSDT.BINANCE USD-M perps that only took a break when open interest was expanding into it (new capital committed), exiting via an ATR chandelier trail, with just three searched parameters.

Over 2187 days and 426 trades (244 long / 182 short) it returned -11.7% with Sharpe -0.038, Sortino -0.030, profit factor 0.960, win rate 40.4% and expectancy -$9.43/trade; avg_trade_return_pct was 0.177% (just above the 0.15% USD-M floor) but commissions were 12.9% of gross and modeled impact was 1762% of gross PnL at capacity_usd ~$322. It was profitable in only 1 of 6 calendar years (2022 +17.4%), with 19.5% max drawdown and a 1368-day drawdown duration.

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