I Re-Created A Quant Trading Strategy With Claude Code (Insanely Cool)

Finance-focused summary (markets / quant methodology)

Core idea: “Hedge fund method” instead of chart indicators / trend lines

• The video claims hedge funds/quants classify market regimes using probabilities derived from historical return transitions, rather than technical indicators or subjective “feelings”.
• Markets are modeled with three regimes (“states”):
  • Bull: last 20 trading days cumulative return ≥ +5%
  • Bear: last 20 trading days cumulative return ≤ −5% (text also implies something like “−6 so on and so forth”)
  • Sideways: anything in between

Framework / step-by-step methodology (10 elements)

1. Define “states” using the 20-day cumulative return thresholds

   • Bull if ≥ +5%
   • Bear if ≤ −5%
   • Sideways otherwise

2. Determine today’s state

   • Label each day in history (starting from day 20) according to the 20-day return criterion.

3. Markov property (focus on today)

   • Transitions depend primarily on the current state, not the full past path.

4. Build the “hedge fund matrix” (3×3 transition matrix)

   • From today’s state → tomorrow’s state, using empirical transition counts converted into probabilities.

5. Persistence / “stickiness”

   • Diagonal elements indicate how likely the regime is to remain the same.
   • Stickiness is inferred from the transition probabilities.

6. Multi-day forecasting by “squaring the matrix”

   • 2-day: square the transition matrix (matrix power 2)
   • 3-day: cube it (power 3)
   • Generally: higher powers for longer horizons; probabilities become less informative as horizon increases.

7. Stationary distribution

   • With very long horizons (example mentioned: ~28 days), probabilities converge toward an uninformative regime mix.

8. Signal generation (probability differential)

   • Trading signal is computed as:
     • Signal = P(tomorrow = Bull) − P(tomorrow = Bear)
   • Interpretation:
     • Magnitude ⇒ trade size / risk intensity
       • “The larger the number, the more money”
     • Sign ⇒ direction
       • Positive ⇒ go long
       • Negative ⇒ go short

9. Walk-forward backtesting

   • Recalculate the entire regime model repeatedly (each day) to avoid “future leakage” from fitting on the whole sample.

10. Hidden Markov Model (HMM) to reduce subjective thresholds

    • Uses pattern recognition to infer state “personalities” rather than relying only on fixed human-labeled bull/bear/sideways thresholds.
    • The video frames overlap between:
      • Original subjective thresholds (±5% over 20 days)
      • HMM-inferred states
    • When they agree, it’s framed as a “green light”.

---

Explicit instruments / tickers mentioned

• Bitcoin
• Ethereum
• XRP
• SPY (used as a demo for the “10-year” / SPY 10-year chart in the onboarding demo)
• Tesla (TSLA)

Also referenced conceptually:

• “Bit Tensor” / “subnets” (example strategy; not clearly tied to a specific ticker in the subtitles)

---

Key numbers / thresholds / example probabilities

State definition thresholds (main parameter)

• Bull if 20-day cumulative return ≥ +5%
• Bear if 20-day cumulative return ≤ −5%
• Otherwise: Sideways

Stickiness / persistence

• Example described (illustrative, not computed live): bull “stickiness” 80%

Signal example (probability differential)

• P(Bull tomorrow) = 65%
• P(Bear tomorrow) = 20%
• P(Sideways tomorrow) = 15%

Signal:

• 65% − 20% = +45% ⇒ long

On-chart example probabilities (Bitcoin, shown in subtitles)

• P(Bull tomorrow) = 29%
• P(Bear tomorrow) = 42%
• P(Sideways tomorrow) = 29%

Interpretation:

• Bear is most likely based on recent behavior/stickiness.

Long-horizon caution

• Example: ~28-day forecast
  • Probabilities converge toward a uniform / low-signal distribution, reducing actionable usefulness.

---

Recommendations / cautions and risk handling

• Backtesting bias caution

  • Fitting once over the full dataset can “learn from the future”.
  • Use walk-forward recalculation to mitigate this.

• Risk sizing concept

  • Trade size scales with the strength of the probability differential (Bull − Bear).

• Horizon warning

  • Longer projections (example: 28 days) may converge toward uninformative distributions, lowering signal quality.

---

Disclosures / disclaimers

• None explicitly shown as a formal “not financial advice” statement in the subtitles.

---

Tools promoted (implementation details)

• A Claude Code / LLM “skill” with a one-shot prompt installed via GitHub.
• TradingView Pine Script
  • Used to visualize the 3×3 probability matrix on charts.
• Setup timing mentioned:
  • Skill install: ~90 seconds (Mac/Linux), 2–3 minutes (Windows)

---

Presenters / sources mentioned

• Rowan
  • Quant reference for the original “hedge fund method” / thread
  • Also referenced again as “Rowan Chain, an observable marov regime model”
• Lewis
  • Speaker name appearing in subtitles (e.g., “Well, Lewis, didn’t we just do…”)