S&P 500 Options (Straddles)
Direct options trading and why equity-style cost models fail for options
Options
Daily
Price Data
Methodology Highlight
Teaches why equity-style cost frameworks are structurally wrong for options, and the critical difference between mid-price backtests and executable-price reality.
Unlike the equity+options case study that uses options data to predict stocks, this case study trades options directly. It sells ATM straddles on S&P 500 constituents and delta-hedges daily, testing whether the variance risk premium — the persistent gap between implied and realized volatility — can be harvested systematically.
The defining lesson is about cost modeling. Option bid-ask spreads are quoted in volatility points and scale with premium, not with notional value. This makes standard basis-point cost sweeps (which work for equities and futures) structurally misleading for options. Students learn to build executable-price backtests that incorporate actual bid-ask spreads, hedge costs, and margin economics.
The case study demonstrates that a statistically strong ML signal does not guarantee a viable strategy — the gap between mid-price and executable- price performance can be enormous in options markets. This teaches the critical distinction between prediction quality and economic viability.
Strategy Summary
Sell ATM straddles on S&P 500 constituents weekly, delta-hedge daily, exit after 10 days. 612 individual equity straddles with 51 IV-derived features. The ML signal is built on delta-hedged returns. Cost model includes option spread, hedge spread, commission, and margin opportunity cost — evaluated at mid-price and executable-price levels.
Data Sources
OptionMetrics (options data)
CBOE (VIX, VVIX)
ML Techniques
IV-derived feature engineering
Delta-hedged return labels
Options-specific cost modeling
Multi-label backtesting
ML Pipeline
Universe & Setup
ATM straddles on S&P 500 constituents. The only case study that trades options
directly. Weekly Friday entry, daily delta-hedge, 10-day hold. A unique
`02_instruments` notebook resolves contract roll contamination: materialized
straddles roll contracts almost daily (68% roll rate), so shift-based forward
returns mix different contracts. Same-contract exit price lookups from raw option
chains produce clean labels. Margin requirements: 15-20% initial.
Universe & Protocol Setup
Ch 6
Defines contract selection rules for ATM straddles on S&P 500 constituents.
Documents P&L accounting for delta-hedged short straddles, hedging protocol
(daily delta rehedge), margin requirements, and the dominant cost structure.
Builds walk-forward CV splits.
Contract Instruments
Ch 6
Resolves a critical data quality issue: materialized straddles roll contracts
frequently, so shift-based forward returns mix different
contracts -- creating phantom returns from chaining contracts with different
time values. Implements same-contract exit price lookups from the raw option
chain to produce clean labels.
Labels & Evaluation
Three label variants enable the defining Sharpe decomposition: unhedged 10-day
(mid-to-mid), delta-hedged 10-day (primary), and executable (at actual bid-ask).
Positive return = profitable short vol position before costs. 51 features (47
financial + 4 temporal) evaluated. The three-label structure reveals that mid-to-mid
Sharpe +2.70 collapses to -1.05 at executable prices.
Label Engineering
Ch 7
Constructs short straddle return labels using same-contract prices from
02_instruments. Implements three label variants: unhedged 10-day (mid-to-mid),
delta-hedged 10-day (primary), and 5-day variants. Positive return means
profitable short vol position before costs. These three labels enable the
Sharpe decomposition that reveals the cost gap.
Feature Evaluation
Ch 7
Evaluates all features (financial + temporal) against delta-hedged straddle
return labels using HAC-adjusted IC. Applies Benjamini-Hochberg FDR
correction. Compares IC against unhedged vs delta-
hedged returns. Assesses feature redundancy and produces triage decisions for
Ch11 modeling.
Feature Engineering
47 features organized into three groups unique to options: instrument state
(straddle mid, Greeks, DTE, relative spread), surface features (IV level, skew,
term structure, VRP, dynamics), and quality features (convergence codes, staleness
proxies). VRP (IV minus RV) with per-symbol normalization. GJR-GARCH temporal
volatility and Bayesian stochastic volatility produce 4 temporal features.
Feature Engineering
Ch 8
Generates features organized into three groups: instrument state (straddle
mid, Greeks, DTE, relative spread), surface features (IV level, skew, term
structure, VRP, dynamics), and quality features (convergence codes, staleness
proxies). Computes VRP (IV minus RV) with per-symbol normalization -- features
unique to options case studies.
Temporal Features
Ch 9
Fits walk-forward volatility models: GJR-GARCH (all symbols) estimates
parameters on training data then filters the full series with fixed parameters
to produce daily conditional volatility -- no re-estimation, no look-ahead.
Bayesian Stochastic Volatility provides a second temporal feature. Produces
temporal features for downstream modeling.
Modeling
The strongest signal environment in the book. LASSO achieves IC 0.215 (richest
linear signal), GBM pushes to IC +0.068, PatchTST reaches IC 0.387 on delta-hedged
returns. All model families positive in both CV folds. A dedicated IC diagnostic
notebook investigates why IC is an order of magnitude above other case studies.
Causal DML tests whether VRP causes higher delta-hedged returns.
Linear Models
Ch 11
Trains Ridge, LASSO, and ElasticNet on IV-derived features across equity
straddles to predict delta-hedged 10-day returns. LASSO achieves the richest
single-family linear signal in the book. Establishes the regularized
regression baseline for downstream GBM and deep learning comparisons.
GBM Grid Search
Ch 12
Trains LightGBM across tree-depth profiles and loss functions on the options
feature matrix. Searches for non-linear interactions among IV skew, term
structure, VRP, and volatility regime features. GBM improves on the linear
baseline on delta-hedged returns, confirming non-linear structure in the
options signal.
Tabular DL (TabM)
Ch 12
Trains TabM rank-1 adapter MLP ensemble (small/medium/large variants) on the
same feature matrix as linear and GBM. Compares attention-based ensembling
against tree-based models for capturing multi-way interactions between IV term
structure, skew, VRP, and cross-sectional positioning.
IC Diagnostic
Ch 11
Decomposes the high IC via four controlled experiments: feature
ablation (IV vs realized-vol features), lag decay (genuine VRP vs same-day
leakage), PCA dimensionality reduction, and cross-sectional breadth analysis.
Separates the VRP prediction component from equity directional signal.
LSTM
Ch 13
Trains LSTM on 60-day lookback windows of IV features across equity
straddles to predict delta-hedged returns. Evaluates whether sequential
processing of volatility clustering and mean-reversion dynamics improves on
flat-feature models (TabM, GBM) that see only the current snapshot.
PatchTST
Ch 13
Trains PatchTST to group 60 days of IV features into patches and apply
self-attention, predicting delta-hedged straddle returns from the evolution
of volatility surfaces rather than their current snapshot. Achieves the
highest single-model IC across all case studies.
Causal DML
Ch 15
Applies Double Machine Learning to estimate the causal effect of vrp_21d on
delta-hedged straddle returns, conditioning on rv_21d, vrp_mom_5d, and
spread_pctl as confounders. Finds the VRP signal is predictive but not causal
-- confounders absorb the entire treatment effect.
Strategy Pipeline
The central tension: strongest IC vs dominant costs. Median round-trip bid-ask spread
of 1,091 bps (10.9% of premium per trade) overwhelms the cross-sectional signal.
Standard basis-point cost sweep shows Sharpe barely changing from 0 to 1,500 bps --
not cost robustness but model irrelevance, because options costs scale with premium,
not notional. Gamma exposure limits and stop-loss protocols tested.
Model Analysis
Ch 11
Compares IC and fold stability across all model families (linear, GBM, TabM,
deep learning, latent factors, causal DML) on delta-hedged straddle returns.
Confirms every family produces positive IC in both CV folds. GBM leads but
single-stock option bid-ask spreads prevent the signal from surviving
execution.
Backtest & Signal Evaluation
Ch 16
Runs plumbing test, parametric sweep across all prediction-signal
combinations, and statistical analysis (DSR, family comparison) for the
straddle strategy. Executes premium-only backtests at mid-price, beginning
the three-layer Sharpe decomposition (mid-unhedged, mid-delta-hedged,
executable) that reveals the cost gap.
Portfolio Construction
Ch 17
Sweeps top predictions x TOP_K concentration x 6 allocators (equal-weight
through MVO) for the straddle strategy at mid-price with margin constraints.
Evaluates whether any allocator or concentration level improves on the
equal-weight baseline given the options-specific return distribution and
embedded leverage.
Transaction Costs
Ch 18
Applies an extended cost grid to the top allocation combinations, revealing
negative executable Sharpe at actual bid-ask prices. Demonstrates that
equity-style bps-of-notional cost models are structurally wrong for options
where costs scale with premium, not notional -- the standard sweep barely
changes Sharpe across the full grid.
Risk Management
Ch 19
Sweeps position-level (stop-loss, trailing stops, time exits) and portfolio-
level (drawdown breakers, daily loss limits) risk controls on the top
allocation combinations. Calibrates trailing stops via MAE/MFE distributions.
Demonstrates that risk overlays cannot rescue a strategy where cost dominance
makes the baseline executable Sharpe negative.
Synthesis & Verdict
The failure that teaches the most. Real ML signal (IC +0.068) completely neutralized
by option bid-ask spreads. Only executable-label backtesting reveals the true burden.
Verdict: Reframe -- the ML signal is a viable input for equity positioning, not a
directly tradable options strategy. Demonstrates why equity-style cost frameworks are
structurally wrong for options.
Strategy Analysis
Ch 20
Synthesizes the three-label Sharpe decomposition (unhedged, delta-hedged,
executable) with the dominant round-trip spread. Produces the "reframe"
verdict: the ML signal is a viable input for equity positioning, not a
directly tradable options strategy. Documents why equity-style cost
frameworks are structurally wrong for options.
Quick Info
Asset Class
Options
Frequency
Daily
Data Type
Price Data
Notebooks
19
Chapters
13
Chapters
6
Strategy Research Framework
7
Defining the Learning Task
8
Financial Feature Engineering
9
Model-Based Feature Extraction
11
The ML Pipeline
12
Advanced Models for Tabular Data
13
Deep Learning for Time Series
15
Causal Machine Learning
16
Strategy Simulation
17
Portfolio Construction
18
Transaction Costs
19
Risk Management
20
Strategy Synthesis