Objective
Test whether GARCH-family conditional volatility can identify useful equity bottom buy entries or improve portfolio risk control, while preserving enough source, model, and execution detail for another researcher to recreate the backtests.
Approach
The campaign mapped primary and authoritative GARCH literature, separated documented volatility-forecasting rules from inferred bottoming overlays, and froze a causal close-only protocol. It implemented deterministic GARCH(1,1), GJR-GARCH(1,1), and EGARCH(1,1) primitives with Normal and Student-t likelihood options, one-step volatility forecasts, drawdown/high-and-declining-volatility entry overlays, and bounded inverse-forecast-volatility sizing.
The state-space sweep used ^GSPC close returns with calibration through 2000-12-29 and holdout from 2001-01-02 through 2026-07-24. Signals were observed at close t and weights shifted at the next close; no holdout selection was used. Fees, one-/two-session execution lag, regime windows, overlapping forward events, fit failures, and data-limited variants were retained explicitly.
Results
The evidence does not support GARCH as a universal directional equity-bottom oracle. Binary high/declining-volatility and drawdown-confirmed entries had positive forward event averages, but lower holdout Sharpe and CAGR than buy-and-hold. Bounded inverse-volatility sizing was the strongest exploratory risk-control overlay: gross GARCH-Normal holdout Sharpe 0.582, CAGR 5.32%, and maximum drawdown -26.67%, versus buy-and-hold 0.449, 7.00%, and -56.78%; at 10 bps it measured 0.522 Sharpe, 4.70% CAGR, and -27.43% drawdown. Its crisis-window Sharpe was negative and its CAGR remained below buy-and-hold.
Five of eight model trials converged, one EGARCH Student-t trial remained an explicit fit failure, and two Realized-GARCH/MS-GARCH or VIX-futures extensions were data-limited by the close-only public cache. No candidate was promoted to the production indicator or strategy catalog. The ranked frontier and next questions remain in research/findings/garch-bottoming-synthesis.json.
Reproduction paths
- Primitive and overlay implementation:
research/garch_bottoming.py - Frozen model and execution spec:
research/findings/specs/garch-bottoming.json - State-space runner and ledger:
research/garch_bottoming_experiment.py,research/findings/garch-bottoming-trial-ledger.json - Diagnostics runner and outputs:
research/garch_bottoming_diagnostics.py,research/findings/garch-bottoming-diagnostics.json - Campaign descriptor and public fixture:
research/campaigns/garch-bottoming/campaign.json,research/fixtures/garch-bottoming.csv - Focused parity and causality tests:
tests/test_garch_bottoming.py - Executed notebook and converted analysis:
research/experiments/1294.ipynb,docs/experiments/1294-analysis.html - Casual-reader interactive article and payload:
docs/research/garch-bottoming.html,docs/site-data/research/garch-bottoming.json
This is exploratory research. Thresholds, drawdown gates, holding periods, and close-only adaptations are inferred test configurations rather than universal source-paper rules. A future promotion decision would require source-faithful rule comparison, point-in-time walk-forward refits, independent investable data, full costs, overlap-aware event inference, and regime-stability evidence.