The AI-driven fund manager outpacing the market
While peers book tee times, Thomas Rice opens his code editor. He's a stock picker who writes his own investment software-and it shows up in his numbers. The edge isn't hype. It's clean data, disciplined models, and risk controls that run like clockwork.
What actually sets his process apart
- Software-first mindset: Treats the portfolio like a product. Versioned models, release notes, telemetry, and kill-switches.
- Data respect: Fundamentals, pricing, alt data-cleaned, labeled, and tested before a single trade.
- Model humility: Ensembles over hero models. Assumes decay. Retrains on a schedule.
- Risk before returns: Controls factor drift, capacity, and drawdowns with the same care as signal strength.
- Execution detail: Slippage, fees, and venue selection measured and monitored daily.
- Feedback loops: Every trade updates the priors. What's measured gets better.
The stack, simplified
- Ingest: Structured fundamentals, pricing, and curated alternative datasets. No "mystery" feeds.
- Features: Rolling fundamentals, quality and sentiment flags, market microstructure signals.
- Training: Walk-forward validation, nested cross-validation, out-of-time tests.
- Risk: Exposure caps (sector, factor, liquidity), scenario tests, and stress across regimes.
- Execution: Smart order routing, participation caps, and real-time slippage tracking.
- Monitoring: Drift detection, model health dashboards, and automatic fallbacks.
Guardrails that protect P&L
- Capacity: Explicit limits by signal confidence and market depth.
- Costs: Every backtest includes trading costs, borrow fees, taxes, and latency assumptions.
- Overfitting checks: Freeze features before tuning. Penalize complexity. Test on cold periods.
- Explainability: Feature attribution for compliance and PM sanity checks, not just for show. See guidance like the NIST AI Risk Management Framework.
- Regime awareness: Signals behave differently during liquidity shocks and earnings seasons. Position sizes adapt.
What "beating the market" looks like here
Less fanfare, more consistency. High hit rate isn't the goal; positive expectancy with tight loss control is. Turnover fits the signal half-life. Position sizes reflect confidence, liquidity, and crowding-not ego.
Apply this in your team over the next quarter
- Week 1: Define one clear edge (e.g., quality + revisions). Write the hypothesis. Decide the kill criteria.
- Week 2: Build a clean data pipeline with unit tests. No modeling yet.
- Week 3: Create features you can explain to a committee in two minutes.
- Week 4: Run walk-forward tests with realistic costs and position limits.
- Week 5: Add risk overlays: sector caps, beta neutralization, liquidity filters.
- Week 6: Paper trade with full monitoring: slippage, drift, attribution.
- Week 7: Go live with small size. Daily post-trade reviews.
- Week 8: Ship improvements weekly. One change at a time. Document everything.
Need a shortcut to vetted tools and training? Explore curated AI tools for finance and practical courses by job role.
Common traps that kill returns
- P-hacking: Too many features, too little theory.
- Ignoring costs: Backtests that forget fees and borrow rates are fiction.
- Alt-data glitter: "Unique" datasets that don't scale or decay fast.
- Over-automation: Humans out. Errors in. Keep a manual override.
- Black-box comfort: If you can't explain the driver, you can't size the risk. For perspective, see research like this SSRN paper on ML limits in asset management.
The takeaway
Rice's advantage isn't luck. It's a repeatable system: clean inputs, modest models, strict risk, fast iteration. You don't need a giant team-just a clear edge, working code, and the discipline to measure what matters.
Your membership also unlocks:
