Skip to content

Latest commit

 

History

History
324 lines (237 loc) · 11.3 KB

File metadata and controls

324 lines (237 loc) · 11.3 KB

Strategy Logic: Why This Model Works

Executive Summary

This strategy exploits information asymmetries and microstructure patterns in the QQQ options market to forecast next-day price movements. The core thesis: options market participants reveal their expectations and hedging needs through pricing, volume, and positioning—creating predictable short-term dynamics.

Core Rationale

1. The Options Market as a Sentiment Barometer

Key Insight: Options are insurance contracts. When market participants buy insurance (puts), they pay a premium. When they sell insurance (puts), they collect premium. These flows reveal true positioning, not just stated opinions.

Why It Matters:

  • Smart money hedges through options (institutions, hedge funds)
  • Retail traders speculate through options (often contrarian signals)
  • Market makers must hedge their books, creating predictable flows

Unlike equity order flow (which can be hidden), options activity is visible through volume, open interest, and implied volatility changes.


Key Predictive Features & Their Financial Logic

1. Variance Risk Premium (VRP)

Definition: VRP = Implied Volatility (IV) - Realized Volatility (RV)

Financial Intuition:

  • IV represents the market's expectation of future volatility
  • RV represents actual historical volatility
  • VRP typically positive: investors pay a premium for insurance (puts)

Predictive Power:

  • High VRP (IV >> RV): Market overpricing risk

    • Insurance is expensive → Mean reversion opportunity
    • Often occurs after selloffs when fear is elevated
    • Trading Signal: Potential long opportunity
  • Low/Negative VRP (IV < RV): Market underpricing risk

    • Insurance is cheap → Suggests complacency
    • Precedes volatility spikes
    • Trading Signal: Reduce exposure or hedge

Academic Support:

  • Carr & Wu (2009): "Variance risk premiums"
  • Bollerslev et al. (2009): VRP predicts equity returns

Our Implementation:

VRP = IV_ATM_Monthly - RV_20d
  • Use 1-month ATM IV (most liquid, representative)
  • Compare to 20-day realized vol (captures recent regime)

2. Gamma Exposure (GEX)

Definition: Total gamma across all options, weighted by open interest and strike.

Financial Intuition: Market makers (MMs) are short options to customers. To remain delta-neutral, MMs must:

  • Buy the underlying when it goes down (supporting prices)
  • Sell the underlying when it goes up (capping prices)

This creates pinning effects around high gamma strikes.

Predictive Power:

  • High Positive GEX: Large dealer hedging needs

    • MMs act as stabilizers (buy dips, sell rips)
    • Price suppression (reduced volatility)
    • Trading Signal: Lower expected movement, fade extremes
  • Low/Negative GEX: Dealers have inverted exposure

    • MMs become momentum amplifiers (sell dips, buy rips)
    • Price acceleration (increased volatility)
    • Trading Signal: Trend-following, risk-off in crisis

Practical Example:

  • If QQQ is at $400 and there's massive call open interest at $400
  • Dealers are short those calls → long gamma
  • As QQQ rises to $405, dealers must sell QQQ to hedge
  • This selling pressure caps the rally

Our Implementation:

GEX = Σ (Gamma × Open_Interest × Strike × 100)
GEX_net = GEX_calls - GEX_puts
  • Focus on ATM gamma (most sensitive)
  • Normalize by notional value to make comparable over time

Industry References:

  • SqueezeMetrics (pioneered GEX research)
  • SpotGamma (dealer positioning analysis)

3. Put/Call Ratios

Definition: Ratio of put volume (or OI) to call volume (or OI).

Financial Intuition: Measures directional sentiment in the options market:

  • High PCR (>1.5): More puts than calls → Fear/hedging
  • Low PCR (<0.7): More calls than puts → Greed/speculation

Predictive Power (Contrarian):

  • Elevated put buying often marks bottoms

    • Excessive fear → Capitulation → Reversal
    • Especially OTM puts (retail panic)
  • Elevated call buying often marks tops

    • Excessive greed → Complacency → Correction
    • Especially OTM calls (speculative fervor)

Our Implementation: We calculate three versions:

  1. PCR_Volume: Put volume / Call volume (intraday sentiment)
  2. PCR_OI: Put OI / Call OI (longer-term positioning)
  3. PCR_OTM: OTM put volume / OTM call volume (speculative activity)

Why Multiple PCRs?

  • Volume = Today's activity (faster signal)
  • OI = Accumulated positions (slower signal)
  • OTM = Speculative extremes (best for reversals)

4. Volatility Skew

Definition: Difference in implied volatility between OTM puts and OTM calls.

Financial Intuition: In equity markets, puts trade at a premium to calls (negative skew) because:

  • Investors hedge downside (buying puts)
  • Downside moves are faster and scarier than upside moves
  • "Crashophobia" since 1987

Predictive Power:

  • Steepening skew (put IVs rising faster than call IVs):

    • Increasing demand for downside protection
    • Trading Signal: Bearish/cautious
  • Flattening skew (call IVs catching up):

    • Decreasing fear, increasing speculation
    • Trading Signal: Bullish/risk-on

Our Implementation:

Vol_Skew_Monthly = IV_OTM_Put_Monthly - IV_OTM_Call_Monthly
Vol_Skew_Deep = IV_Deep_OTM_Put_Monthly - IV_Deep_OTM_Call_Monthly
  • Track both OTM (near the money) and deep OTM (tail risk)
  • Deep OTM skew = Tail risk pricing (black swan hedging)

Academic Support:

  • Bates (1991): "Crash of '87: Was it expected?"
  • Rubinstein (1994): "Implied binomial trees"

5. Momentum & Realized Volatility

Definition:

  • Momentum: Cumulative returns over X days
  • Realized Vol (RV): Standard deviation of returns (annualized)

Financial Intuition:

  • Momentum: Captures trend strength (persistence)
  • RV: Captures market regime (calm vs. chaotic)

Predictive Power:

  • High RV: Market stress

    • Higher transaction costs, wider spreads
    • Mean reversion more likely (overshoots)
    • Trading Signal: Reduce leverage, fade extremes
  • Low RV: Market calm

    • Trends persist longer
    • Momentum strategies work better
    • Trading Signal: Increase leverage, follow trends

Our Implementation:

  • Calculate RV over 5d, 10d, 20d, 60d windows
  • Calculate momentum over same windows
  • Allows model to detect both short-term and long-term regimes

Interaction Effects

The model captures non-linear interactions between features:

1. GEX × Momentum

  • High GEX + Positive Momentum: Price pinning (resistance)
  • Low GEX + Positive Momentum: Price acceleration (breakout)

2. VRP × Skew

  • High VRP + Steep Skew: Maximum fear (contrarian buy)
  • Low VRP + Flat Skew: Complacency (risk-off)

3. PCR × RV

  • High PCR + High RV: Panic selling (reversal signal)
  • High PCR + Low RV: Rational hedging (no reversal)

Regime Detection

Why It Matters: Strategies that work in low volatility (trend-following) fail in high volatility (mean reversion).

Our Approach:

  • Define regimes based on 20-day RV vs. 60-day MA:
    • High Vol Regime: RV_20d > MA_60(RV_20d)
    • Low Vol Regime: RV_20d ≤ MA_60(RV_20d)

Adaptive Strategy:

  • Low Vol: Full leverage, momentum-oriented
  • High Vol: Reduced leverage, mean reversion
  • Extreme Vol (>50%): Go to cash (crisis mode)

Why Ensemble Models?

Single Model Problem:

  • Overfitting: Captures noise, not signal
  • Instability: Small data changes → big prediction changes

The "High Calmar" Trap: During the R&D phase, we experimented with advanced Deep Learning architectures, including LSTMs and Transformer Encoders, and aggressive hyperparameter optimization using Optuna, to capture sequential dependencies in the volatility surface. These models achieved spectacular in-sample performance, with one Transformer variant reaching a Calmar Ratio of 4.7. They lacked robustness. A 10% change in lookback windows caused performance to collapse. The models were "memorizing" the noise of the specific training period rather than learning structural market mechanics.

Ensemble Solution: Combine 4 diverse models:

  1. LightGBM (30%): Fast, captures complex interactions
  2. XGBoost (30%): Robust, regularized
  3. Random Forest (30%): Bagging reduces variance
  4. Ridge Regression (10%): Linear anchor (prevents overfitting)

Why This Works:

  • Boosting (LGB/XGB) captures non-linear patterns
  • Bagging (RF) provides stability
  • Linear (Ridge) prevents overfitting to noise
  • Equal weights (mostly) = Simple, robust

Risk Management Framework

1. Volatility Targeting

Position_Size = Target_Vol / Realized_Vol
  • Scale down during high vol (risk control)
  • Scale up during low vol (opportunity)
  • Capped at 1.5x (regulatory/broker limits)

2. EMA Smoothing

Signal_t = α × Prediction_t + (1 - α) × Signal_{t-1}
  • Reduces churn (fewer trades, lower costs)
  • Prevents whipsaws (false signals)
  • α = 0.15 (empirically optimal)

3. Regime Kill Switch

If RV > 50%: Position = 0
If RV > 35%: Position *= 0.5
  • Protects during black swan events
  • Prevents catastrophic drawdowns
  • Historical examples: March 2020, Oct 2008

Hypothesis Testing: Why Next-Day?

Why predict 1-day ahead (not 1-week or 1-month)?

  1. Information Decay: Options flows are high-frequency signals

    • Today's unusual call buying is relevant today/tomorrow
    • By next week, information is stale
  2. Market Efficiency: Markets are more efficient over long horizons

    • 1-day inefficiencies exist (microstructure, dealer hedging)
    • 1-month inefficiencies are arbitraged away
  3. Compounding: Small daily edge compounds

    • 0.1% daily edge = 25% annual return
    • Sharpe of 2+ over 252 days

Limitations & Failure Modes

When This Strategy Struggles:

  1. Flash Crashes: Model can't predict exogenous shocks (e.g., COVID-19 announcement)
  2. Regime Shifts: Takes ~20 days to detect new regime (lagging indicator)
  3. Low Liquidity: Options data unreliable during market closures or holidays
  4. Structural Changes: Fed policy shifts, new market participants (retail surge 2020-2021)

Conclusion

This strategy succeeds because it:

  1. Exploits real market mechanics (dealer hedging, sentiment, insurance pricing)
  2. Uses ensemble learning (reduces overfitting, increases stability)
  3. Adapts to regimes (different strategies for different markets)
  4. Manages risk (volatility targeting, kill switches)

The options market is a noisy but informative signal. By carefully engineering features, applying robust machine learning, and managing risk, we extract a consistent edge.


References:

  • Carr, P., & Wu, L. (2009). Variance risk premiums. The Review of Financial Studies, 22(3), 1311-1341.
  • Bollerslev, T., Tauchen, G., & Zhou, H. (2009). Expected stock returns and variance risk premia. The Review of Financial Studies, 22(11), 4463-4492.
  • Bates, D. S. (1991). The crash of'87: Was it expected? The evidence from options markets. The Journal of Finance, 46(3), 1009-1044.
  • Gârleanu, N., Pedersen, L. H., & Poteshman, A. M. (2009). Demand-based option pricing. The Review of Financial Studies, 22(10), 4259-4299.