Video summary

I Re-Built A Quant Trading Strategy With Fable 5

Main summary

Key takeaways

Finance

Finance-specific summary (quant/hedge-fund “Markov” trading framework)

The video describes a “hedge fund method” for quant trading that replaces subjective chart indicators with a regime-based approach. It uses:

  • State classification of market conditions
  • Transition probabilities between states to generate trade direction and (optionally) position sizing.

Instruments / tickers / assets mentioned

  • SPY (S&P 500 ETF)
  • Bitcoin (BTC) (referenced via “Bitcoin chart”; ticker not explicitly stated)
  • Ethereum (ETH) (referenced; ticker not explicitly stated)
  • Tesla (TSLA) (referenced; ticker not explicitly stated)
  • S&P 500 (index)
  • “S&P 500” backtest over ~30 years (mentioned as a prior prompt result)

No other tickers (e.g., bonds, sector ETFs, commodities) are mentioned.

Key numbers / thresholds / outputs mentioned

  • State lookback window: last 20 days of price history

  • Example regime thresholds for returns over 20 days:

    • Bull state: return > +5%
    • Sideways state: return between -5% and +5%
    • Bear state: return < -5%

  • “Stickiness” fix (methodology change):

    • Instead of comparing every day with overlapping 20-day windows, the method waits for non-overlap (i.e., uses a new 20-day segment without overlap with the prior segment) to avoid inflated stickiness scores.

  • Confidence / signal logic:

    • Compute likelihood of bull next day and bear next day
    • Signal strength uses: (bull likelihood − bear likelihood)
    • Example: 80% bull vs 10% bear → 70% net bullish confidence

  • Multi-day projection:

    • If 1-day bull likelihood is 60% (0.6), then 2-day likelihood uses multiplication: 0.6 × 0.6 = 0.36 (36%)

    • General idea: cube for 3 days, power-of-n for n days

    • Warning: far-ahead projections become impractically small.

  • Walking forward / leakage correction:

    • Prior version reportedly made S&P 500 lose money over ~30 years
    • Improved “no future leakage” version shows S&P 500 profitable (stated as reality)
    • Bitcoin example: previously claimed 23x growth, improved to nearly 60x growth

  • Live demo snapshot (Bitcoin):

    • “Today is a bear state”
    • BTC: “minus 16.7% over the last 20 bars” on the daily chart (as shown in the demo)

Methodology / step-by-step framework (as described)

  1. Define market “states” (regimes) by quantifying market direction over a lookback window.
  2. Classify state using return over the last 20 days into:
    • Bull / Sideways / Bear (example thresholds: >+5%, between -5% and +5%, < -5%)
  3. Apply the Markov property:
    • Next-day state depends primarily on the current state, not the detailed past path.
  4. Compute transition probabilities:
    • Back over history; count transitions between states (e.g., bull→sideways, bull→bear, etc.)
    • “Nine combinations” are referenced (implying 3 states with pairwise transitions).
  5. Measure “stickiness”:
    • Bull likely to remain bull; bear likely to remain bear.
    • Improvement: recalculate using non-overlapping 20-day windows to reduce inflated stickiness.
  6. Create a trading signal from state probabilities:
    • Use: (P(bull next day) − P(bear next day))
    • If bullish net is positive → long bias (as implied)
  7. Position sizing / scaling by confidence (recommended concept):
    • Allocate more when confidence is high (e.g., ~70% net bullish in the example)
    • Allocate less when confidence is low
  8. Project further ahead (optional):
    • Use repeated multiplication/powers of probabilities for 2–3+ days
    • Warns very far ahead becomes less useful due to tiny numbers
  9. Walking forward backtesting (risk/validity fix):
    • Backtests must not “see the future.”
    • Restrict training/data access to only information before the evaluation window to avoid “exam after seeing the answer sheet.”
  10. Hidden Markov method (parameter-free state labeling):
    • Instead of fixed ±5% boundaries, infer regime boundaries from the data.
    • Use inferred labels to reduce arbitrariness.
  11. Generate / installation workflow:
    • Use an LLM prompt to create a Pine Script for TradingView that computes regimes, transitions, stickiness, and emits a current signal.

Explicit recommendations / cautions

  • Caution on backtest validity: naive backtesting can “learn from events ahead” (data leakage). The video frames “walking forward” as critical.
  • Caution on stickiness estimation: overlapping windows can inflate stickiness; recalculate using no overlap.
  • Boundaries shouldn’t be arbitrary: fixed thresholds (like ±5%) are described as “made up.” “Hidden Markov” is presented as the fix.
  • Confidence-based trading: scale exposure by probability/confidence rather than going all-in on every signal.
  • Far-ahead forecasting likely not useful: probabilities shrink toward zero over multiple steps.

Note: No explicit “not financial advice” disclaimer is included in the subtitles provided.

Presenter / sources mentioned

  • Roan (Twitter/X quant credited with the original method; work condensed)
  • Fable 5 (LLM model mentioned for improving the prompt/framework; previously also Opus 4.7)
  • Claude (LLM used in the copy/paste workflow)
  • TradingView (platform receiving the Pine Script / chart connection)
  • 01 Systems (classroom/prompts link/source for the prompt)

Original video