Back to content
03AI for Builders

I Built an AI Trading Desk (And You Can Too)

What if you could build your own trading desk — covering 676 assets, running five strategies, with AI agents optimizing overnight — for less than a Bloomberg terminal subscription? I did it in 8 days.

35-38 min
AI TradingQuantitative FinanceAsset RotationAI AgentsBuilding Systems

Listen Now

Spotify

Episode Summary

⚠️ IMPORTANT DISCLAIMER: I am not a financial advisor. This is experimental, educational content about a personal project. Do not trade based on anything I say. You can lose money. Consult a licensed professional.

AI is going to replace my job as a test engineer in 5-10 years. While I still have stable income, I'm building something to replace that income. This episode covers the MRE (Multi-Regime Engine) — a systematic trading system that passed JP Morgan's institutional analysis framework.

The Why: The Clock Is Ticking

The reality: AI will eliminate most test engineering roles within the decade. The choice is to either bury my head in the sand or use this window to build something sustainable.

The problem with buy-and-hold: Traditional 60/40 portfolios assume infinite time and emotional resilience. I have neither.

  • 2000-2013: If you bought the S&P at the dot-com peak, you didn't break even until 13 years later
  • 2022: Both stocks AND bonds fell simultaneously — diversification failed
  • Drawdowns of 30-50% can take years to recover from

What I need: $200K annual income from trading to sustain my family when my engineering salary disappears.

The Solution: Asset Revesting

Inspired by Chris Vermeulen's research, the MRE uses regime-aware asset rotation instead of buy-and-hold:

  • Bull Regime: Buy risk assets (stocks, growth)
  • Bear Regime: Rotate to defensive assets (bonds, cash, gold)
  • Sideways Regime: Wait for clarity, hold cash

The insight: You don't need to predict the future. You just need to follow the regime.

The MRE System Architecture

Five Independent Strategies

  1. Fear and Greed Contrarian — Buy extreme fear (CNN index below 30), sell extreme greed (above 70)

  2. Regime Confirmation — Buy dips within established bull regimes (2-5% pullbacks)

  3. RSI Oversold — Buy when Relative Strength Index drops below 30 (mean reversion)

  4. Bollinger Bands Mean Reversion — Buy when price touches lower band

  5. Pair Mean Reversion — Trade convergence of 34,760 correlated asset pairs

The Data Stack

  • 676 tickers analyzed — Full S&P 500, NASDAQ-100, Dow 30, plus 159 ETFs
  • Regime signals for each asset with exponential moving averages
  • Fibonacci retracements and projections for entry/exit zones
  • VIX data — 30 years of volatility history for crash detection
  • Market breadth — RSP/SPY ratio and internal breadth metrics
  • Prediction markets — Kalshi and Polymarket data for macro events
  • Sector rotation tracking across 10 major sectors

The Agent Fleet

Four AI agents manage the system:

  1. The Updater — Refreshes prices every 5 minutes during market hours

  2. The Advisor — Adds qualitative analysis, can override quantitative signals

  3. The Analyst — Reviews daily performance, tracks accuracy by strategy

  4. The Overnight Optimizer — The Beast

    • Five-phase optimization pipeline
    • 300,000+ backtests across all parameters
    • Walk-forward validation (no overfitting)
    • Monte Carlo robustness testing
    • 39 minutes to optimize all 676 tickers

JP Morgan Validation Results

We tested the system through JP Morgan's institutional framework:

Statistical Performance

  • Over 4,100 backtests across 27 configurations
  • 60.09% overall accuracy, 65.65% buy signal accuracy
  • Sharpe ratio: 1.13 (anything above 1.0 is good)
  • P-value essentially zero — statistically significant results

Institutional Capacity

  • $100-500 million deployable according to JP Morgan's analysis
  • APPROVED FOR DEPLOYMENT with tier-1 liquidity
  • MEDIUM-HIGH sustainability with 2-3 year alpha half-life

Key Differentiators

  • 96.6% accuracy from Kalshi prediction market integration
  • Multi-asset confirmation signals
  • Regime-aware position sizing
  • Cross-asset correlation insights

Today's Plays: From Signals to Action

The system synthesizes 676 tickers × 5 strategies into four actionable categories:

  • BUY — Strong signal, bull regime, multiple strategy confirmation
  • WATCH — Sideways regime, waiting for breakout
  • HOLD — Bull regime but price has run too far
  • WAIT — Bear regime, capital preservation mode

Position sizing: Higher confidence = larger allocation (8% at 50% confidence, 25% at 90%)

Paper Trading Reality Check

Current performance: Started February 9th with $100K, now at $99,340 (down 6.45%)

Eight open positions: SPY, EFA, GLD, TLT, VNQ, XLE, XLF, XLV

Why I'm sharing this: Building in public means showing the drawdowns too. The question isn't whether you have bad weeks — it's whether the process is sound and drawdowns are within expected parameters.

The Bug War Stories

Three Critical Bugs That Almost Broke Everything:

  1. The Five-Minute Hold — System counted hold days in poll cycles instead of calendar days, executing 50 rapid trades and losing $5,356

  2. The Rogue Dashboard — UI was generating its own conflicting signals, overriding the main engine

  3. Cross-Strategy Conflicts — Multiple strategies trying to trade the same asset with different parameters

The lesson: Real building is messy. The gap between architecture diagrams and live systems is where the learning happens.

Faith Connection

"The plans of the diligent lead surely to abundance, but everyone who is hasty comes only to poverty."
— Proverbs 21:5

Diligence over talent. This system is about:

  • Showing up every day
  • Building processes that work while you sleep
  • Running 300,000 backtests because you want to get it right
  • Removing emotion and trusting the process

Thirteen versions in eight days. Not because I'm talented — because I kept showing up.

Key Takeaways

  1. Buy-and-hold isn't always safe — Asset rotation can protect during downturns and capture upside during recoveries

  2. Regime signals cut through noise — Don't react to headlines. Read the underlying trend.

  3. One strategy isn't enough — Diversify your signals like you diversify assets

  4. Validate with institutional frameworks — Don't just guess. Prove it works.

  5. Bugs are features in disguise — Every failure made the system stronger

  6. AI agents can manage complexity — A quant team for $50/month in API calls

  7. Be honest about performance — Show the drawdowns, not just the wins

Today's Plays Newsletter

Want to follow this experiment in real-time? I publish the MRE system's daily signals in "Today's Plays" newsletter:

  • Daily BUY/HOLD/WATCH/WAIT signals across 676 tickers
  • Confidence levels and Fibonacci entry zones
  • Real performance tracking — wins AND losses
  • Architecture details and agent configurations

Link in show notes. It's free.

Episode Quote

"Work done in secret has its own reward. The agents run whether I'm watching or not."

🎧 Listen Now

Subscribe wherever you listen to podcasts. New episodes weekly.

The Builder's Frequency

Weekly insights on engineering, building products, and thinking in decades.

Back to all content