Strategies

GDX–GLD Bayesian Spread Reversion with Online Stability Monitoring

Outcome: Abandoned

GdxGldBayesianSpreadReversion

Outcome Summary

GdxGldBayesianSpreadReversion attempted GDX–GLD convergence trades gated by online Bayesian stability monitoring. After three iterations, it failed Layer 1.5 hypothesis/config consistency verification, which flagged the optimization plan’s fixed 1-DAY bar interval. It was abandoned before backtesting or later evaluation stages, leaving no performance metrics.

Strategy report

Hypothesis

Derived by the Research Lead from Triantafyllopoulos and Montana's Dynamic modeling of mean-reverting spreads for statistical arbitrage: trade deviations in the GDX–GLD price spread only while an online Bayesian TVAR(1) model supports local mean reversion. Use synchronized daily bars on GDX.USEQ and GLD.USEQ, preserving the paper's ETF example. The paper supplies the modeling mechanism, not a validated executable strategy; entry thresholds, exits, sizing and trading parameters below are Research Lead choices. Section 4.5's stated position direction appears inconsistent with convergence of a spread defined as primary minus hedge: this operationalization explicitly shorts a high spread and buys a low spread. Orders execute at the synchronized signal session's close; next-session-open execution is neither requested nor assumed. All performance expectations are unverified. Only one hypothesis is proposed because adaptive estimation and stability monitoring are components of the same trading mechanism.

The source's own ETF example avoids substituting a different market merely to satisfy venue quotas. Paired long/short exposure addresses the portfolio's strongest directional imbalance: 80.7% long-only. Daily equity data supports a longer validation span and lower modeled trading costs than the failed fast crypto spread strategies. Corpus concentration is empty, so instrument-specific experiment shares cannot be assessed. This proposal tests uncertainty-qualified local stability rather than assuming a permanently stationary spread or selecting pairs retrospectively.

verification_loop: Verification failed (Layer 1.5 — hypothesis/config consistency) [class=hypothesis_mismatch]: - optimization_plan fixes bar_interval at 1-DAY

Implementation

Trades synchronized GDX–GLD daily spread deviations using a frozen initialization-only OLS hedge and online Bayesian TVAR(1) stability probability. Buys low spreads and shorts high spreads through opposite-side whole-share ETF legs. Entry risk, leg exposure and gross exposure remain capped; exits cover convergence, instability, adverse spread moves and holding duration.

Preserves the previous implementation's imports, signal estimation, alignment, sizing and exits. Sets hedge_lookback=126, entry_z=1.5, stability_confidence=0.90 and min_capture_fraction=0.004, the least restrictive permitted initialization and entry settings. Adds entry_block_reason as a persisted scalar identifying the actual blocking gate without altering the model. The fixed 504-session model burn-in remains mandatory: with 126 hedge-fit sessions, readiness first becomes possible on synchronized session 630, leaving only 16 potentially eligible sessions in a 645-session run with no prior warmup. The feedback does not identify which gate blocked those sessions, and does not establish absent economic edge. Engineering should supply the plan's 1008 synchronized warmup bars before evaluation, using only data before the locked final-exam boundary; no trade count is guaranteed by this revision. No independent verification or market-data reads were performed. The requested outbox file could not be written because this session's filesystem permissions are read-only; the complete bound JSON is returned here.

Verification Results

Verification failed (Layer 1.5 — hypothesis/config consistency) [class=hypothesis_mismatch]: - optimization_plan fixes bar_interval at 1-DAY

Iteration History

Verification failed (Layer 2 — synthetic scenarios): Parameters used: ['phi_a', 'phi_b', 'stop_z', 'delta_a', 'delta_b', 'entry_z', 'prior_d', 'prior_n', 'leverage', 'prior_mean', 'bar_interval', 'min_notional', 'risk_fraction', 'hedge_lookback', 'max_leg_fraction', 'max_hold_sessions', 'quantity_rounding', '_optimization_plan', 'hedge_refit_policy', 'max_gross_fraction', 'min_capture_fraction', 'same_session_reentry', 'stability_confidence', 'model_burnin_sessions', 'prior_covariance_scale', 'open_basket_rebalancing'] Check that __init__ sets all attributes from self.parameters.get(). - flat_ranging: ValueError: initialization_failed: GDX-GLD hedge does not support opposite-side exposure (bar timestamp: 1735704660000) - volatility_spike: ValueError: initialization_failed: GDX-GLD hedge does not support opposite-side exposure (bar timestamp: 1735704660000)

Verification failed (Layer 3 — sandbox backtest): No trades produced Bar type used: GDX.USEQ-1-DAY-LAST-EXTERNAL, Bars processed: 645 Diagnostics: should_enter() returned a side 0 times over 645 evaluated bars -> your ENTRY CONDITION never triggered. Loosen the entry logic / thresholds. Ensure your strategy produces trades with the given data and parameters.

Outcome Summary

Keep the strategy configuration consistent with the hypothesis’s fixed daily bar interval.

After three iterations, the strategy was abandoned following Layer 1.5 hypothesis/config consistency failure: the optimization plan fixed bar_interval at 1-DAY.

Trade GDX–GLD spread deviations on synchronized daily bars while an online Bayesian TVAR(1) model supported local mean reversion.

No backtest metrics were available; backtesting, optimization, analyst review and risk review were not reached.

Analysis

Dynamic modeling of mean-reverting spreads for statistical arbitrage

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