tr8

Objective: Build a systematic sector rotation strategy where LLM news sentiment is the primary signal and technical conditions act as an entry gate — not the other way around. Validate it with walk-forward methodology before any live capital goes near it.

Design factors: Retail broker constraints were a constant — Alpaca API, PDT rules, fractional shares, paper trading before live. The system needed to respect regulatory constraints without custom broker integrations or exchange-level access.

Infrastructure cost had to stay near zero. The system runs once a day at market close and is dormant otherwise. An always-on server for a daily cron job is the wrong shape.

Validation rigor was non-negotiable. In-sample optimization produces parameters that look good in backtests and fail in production. Walk-forward expanding windows are the minimum standard. Consistency across regimes matters more than peak Sharpe on any single fold.

Technology evaluations and selections:

DuckDB over PostgreSQL or SQLite — columnar storage for time-series financial data, analytical query performance without a server process. Everything lands in DuckDB: signals, trades, decisions, tax records, backtest results.

Modal over always-on VPS — serverless, daily cron pattern, ~$3/month actual spend. No capacity planning, no idle cost, no babysitting.

Gaussian HMM over rule-based regime detection — a 5-state model trained on 8 macro features (VIX, EPU, yield curve, credit spread, Ken French momentum factor, cross-sector volatility, mean pairwise correlation, momentum dispersion) captures regime transitions that hand-coded rules miss. Hard rules still override for extremes: VIX ≥ 30 immediately → HIGH_VOLATILITY regardless of model state. Model uncertainty is highest exactly when you need certainty most — hard rules handle the tails.

Optuna on a Pareto front of Sharpe and max drawdown, penalized by fold variance — optimizing Sharpe alone finds parameters that exploit one favorable regime and collapse on the rest. The variance penalty across walk-forward folds forces parameters that work consistently. Consistent mediocrity outperforms inconsistent brilliance over a full market cycle.

LLM grounding constraint — the prompt instructs the model to reason only from the provided text, not prior training knowledge about a sector’s historical reputation. The signal has to come from current news flow, not from the model’s priors about tech or energy.

Signal construction: up to 20 text excerpts per sector assembled into a grounded prompt. Structured JSON output: sentiment, confidence, bull_strength, bear_strength. Confidence adjusted contrastively: adjusted = confidence × (1 + 0.20 × aligning_strength) × (1 − 0.30 × opposing_strength). High bear_strength on a bullish call reduces confidence by up to 30%.

Regime weight multiplies into the composite score. Shock logic (EPU > 300) separately scales defensives ×1.5 and cyclicals ×0.5 — independent of the HMM state, because news-driven macro shocks can precede HMM state transitions.

Stop hierarchy evaluated in order: circuit breaker (−15% from entry, immediate), sentiment reversal (LLM scores bearish ≥ 0.65), trailing stop (activates after 1× ATR gain, trails at 1.5× ATR below peak), fixed ATR stop (2× ATR below entry until trailing activates). Exit checks every 15 minutes intraday.

An LLM generates a ~200-word narrative after every daily run — what happened, why, what the model saw — stored in DuckDB and surfaced in the dashboard alongside equity curve, sector P&L heatmap, and a “is it worth it?” panel comparing tr8 returns to VTI, SPY, and equal-weight after tax drag.

Repo → — access granted on request