# Volatility Earnings Date Trading Strategy with XGBoost

## Summary

This project designs and evaluates a systematic trading strategy that capitalizes on predictable patterns in implied volatility (IV) around earnings announcements using machine learning (ML). Our main tool is an **XGBoost classifier**, trained to forecast short-term IV spikes, which we convert into long or short straddle options trades depending on the predicted direction.

The strategy focuses on:

- Pre-earnings IV buildup ("volatility spike") and post-earnings collapse ("volatility crush")
- High gamma/vega exposure via ATM straddles with short expirations (5-45 days)
- Model-driven signal generation
- Risk management using stop-loss, take-profit, position caps, and exit-before-earnings

## Thesis

Earnings announcements create predictable mispricings in options markets due to IV behavior. We hypothesize that:

- *"Short-term IV changes around earnings are predictable using engineered features and ML models."*

Instead of separate models for long vs short trades, we use a single probabilistic model to determine the most likely profitable straddle direction (long or short).

## Strategy Pipeline

1. **Data Collection:** Option chains + earnings dates + Tiingo sentiment
2. **Feature Engineering:** IV trend, RV/IV ratio, days to earnings, Greeks, volume
3. **Model Training:** XGBoost with rolling window, weekly retraining, confidence outputs
4. **Signal Generation:** Predict probability of next-day IV expansion
5. **Trade Selection:**
- Long Straddle if p > 0.8
- Short Straddle if 0.6 < p < 0.8 and IV percentile > 70%
6. **Asset Ranking:** Multi-factor score using jump/fall size, RV/IV, theta, vega, and straddle price
7. **Filters:** Liquidity (open interest, spread < 5%), near-ATM, valid expiry

## Risk Management

- Max per trade: 6% of capital, capped at $100K
- Position sizing: Based on model confidence
- Ticker exposure: Max 20%, max 5 active trades
- Exits:
- Stop-loss: −15%
- Take-profit: +30%
- Forced exit: 1 day before earnings

## Results:

1. Starting Cash $10,000,000
2. Backtest Period:
- In-sample (training period): 1/1/2018 - 1/1/2022

- Out-of-Sample Period A: 10/1/2024 – 3/15/2025
- Out-of-Sample Period B: 1/1/2022 – 1/1/2023
- Out-of-Sample Period C: 7/1/2023 - 7/1/2024
- Teams will enter trading competition from 5/7/2025 - 9/7/2025 ("Competition Period") -- top 2 teams will get a prize
- Bugs = Disqualification
3. Live Paper Trading (LPT) Period: **at least one week before final submission date**
- (waived, period too tight for this strategy that eye on earnings 25 days prior)

### Output Requirements:
5. Max Drawdown < 10% for all testing and trading periods
6. Drawdown Duration < 4 months for all periods
7. Sharpe Ratio >= 1 for all periods
8. Daily PnL volatility < 5% of account equity for all periods
9. Stresstesting:
- Monthly return during Covid 2/1/2020 - 5/1/2020 > -15%
- Summer Volatility Spike 7/10/2024 - 8/20/2024 > -15%
- Tariff 3/8/2025 - 4/8/2025 > -15%
- **No entry points were strong enough during this period**

**Use the same set of model parameters to run all these testing periods**

## Bottom Line

Our strategy takes ML predictions with real-world execution by integrating forecasted IV changes into disciplined, risk-aware trades. In-sample performance was strong, but results were mixed when tested Out of Sample, mainly because of the market noise during that period. Even so, the system handled stress well and showed that combining machine learning, options strategies, and strong risk rules can lead to smart trading around earnings announcements.

* *OOS C did well because Unlike early 2022 or early 2025, this period lacked macro shocks [e.g., Fed pivots, wars, inflation scares], allowing option pricing to be driven more by earnings than market noise*

## Future Improvements

- Adjust decision thresholds based on market conditions (adaptive thresholding)
- Identify and respond to different market volatility patterns (regime classification through VIX and Fed rate changes)
- Regularly retrain the model to keep it up to date with recent data

Overall, this project shows that even in unpredictable markets, there’s structure we can learn from, and with the right tools, volatility itself can become a source of profit.

## Alternative Strategy - Adaptive Retraining Model

Instead of separating the data into an in-sample (IS) and out-of-sample (OOS) phase like traditional backtests, this strategy re-trains the model using only the most recent 100 days of feature data. This model adapts to changing market regimes -- like sudden volatility shifts, macro shocks, or sentiment changes.

- This is like running a rolling real-time strategy that constantly re-learns based on recent market behavior, making it adaptable to changes and sudden news cycles


## Metrics for Backtest:

- Sharp Ratio
- Portfolio Returns
- Risks
- Hit Ratio

## Metrics for Risk Management:

- Stop-loss
- Drawdown Limits
- Daily loss limits
- VaR Limits

## Core Requirements:
- [x] Code Clarity
- [x] Articulate of why it may work well in live trading environment (even if we dont have enough sample data yet)
- [x] Quality of Project Research Paper and PPT
- [x] Originality of Core ideas underlying the strategy

## Strategy - ATM Implied Volatility Earnings Date Arbitrage

TLDR:

* **Long Straddle:** Buy Call + Put when it's "cheap", sell it a closer before earnings date when Implied Volatility is higher
* **Short Straddle:** Sell Call + Put when it's "expensive", sell it after earnings date when Implied Volatility is lower

These are At-The-Money calls with minimal skews and straddles are centered.

### Example on QC Research: CSCO


### Considerations

- [x] What are our indicators? i.e. how do we determine if Implied Volatility is "cheap" or "expensive"
- [x] Lookback for historical volatility (One month or quarter? Equal or exponential weight and why?)
- [x] What is the exit strategy? (When IV jumps a certain %? Or when we "deem" the IV to be reaching maximum based on historical performance)
- [x] Gamma/Delta Hedged? Any specific requirements (e.g. keep Delta < 10)
- [x] Other Risk Control Parameters such as stop-loss, position size, VaR
- [x] Best kind of stocks/ETFs to be used for our strategy (stock-choosing rationale)

## Relevant Readings

### [Volatility Trade Design (Louis H. Ederington and J. Scott Chaput)](https://papers.ssrn.com/sol3/papers.cfm?abstract_id=313980)

- The paper argued that straddles, due to their construction, inherently possess high vega (sensitivity to volatility changes) and gamma, making them effective tools for volatility trading.
- **High return + high risk**
- Emphasizing the need for careful timing and risk management when using straddle strategies

### [Anticipating Uncertainty: Straddles around Earnings Announcements (TsingHua)](https://eng.pbcsf.tsinghua.edu.cn/__local/A/11/35/B0A1714D795BD9799087BCCC2AD_A01C4283_54641.pdf?e=.pdf)

This paper argued that most profit comes **before** the announcement, not after.

- IV gradually rises starting about 30 days before
- Becomes steeper 4 days before earnings
- Peak at one day before earnings announcement
- Crash on one day after earnings announcement

**The Paper's Straddle Strategy:**
- Moneyness at 0.9 to 1.1
- Holding period ranges from 1 to 4 days

### [Trading Volatility by Marco Avallaneta](https://www.scribd.com/document/321733308/Harcourt)

- A long vol position (straddle) profits if realized volatility > implied volatility over the holding period​
- Meaning for long straddle, we ideally want RV/IV > 1 or as close to that as possible

### [Volatility Arbitrages by Marco Avellaneda](https://www.scribd.com/document/66344188/Volatility-Stat-Arb-Marco-Avellaneda)

- Paper argued that if we don't manage theta decay carefully (especially in low-vol environments), it can damage our return significantly, thus theta management is critical
- Paper also stressed that **hedging** is the key to maintain market neutrality and get desired results

### [Profitability Analysis of the Straddle Strategy in Trading One-Month Options (Westcliff)](https://wijar.westcliff.edu/wp-content/uploads/2022/12/Cheffa-Shamsa.pdf)

- One-month straddles performed better than two-and-a-half-month straddles in previous studies
- They argued that the most important consideration when doing a straddle is the initial premium
- Overtime, the the ratio of final price to starting price is statistically neutral (supports "random walk" theory for AAPL)
- Profitable exits happen before expiry (so set a profit limit and don't always hold until expiration)

### [Anticipating Uncertainty: Straddles Around Earnings Announcements (Yuhang Xing and Xiaoyan Zhang)](https://quantpedia.com/www/Anticipating_Uncertainty-Straddles_Around_Earnings_Announcements.pdf)

- When VIX is high, straddle returns are lower
- Options (and straddles) tend to be already expensive when overall market fear (VIX) is high
- Be cautious about long straddles when VIX is elevated, prefer short volatility trades
- When VIX is low, options underprice upcoming uncertainty (especially around earnings)
- Long straddles become more attractive because volatility (IV) has more room to spike

### Section 7.4 of "151 Trading Strategies"

- Implied volatility tends to be higher than realized volatility most of the time, which is known as the “volatility risk premium”
- Options are priced higher than what actually happened ("realized volatility")
- Gamma hedging to keep the strategy delta neutral, and it becomes a "theta play", or capitalizing theta decay from selling options
- Catch: Gamma hedge becomes more expensive as the underlying move away from the strike of the sold put & call, if it exceeded option premia, the strategy starts losing money

### Section 7.5 of "151 Trading Strategies"

- OTM put tend to be priced higher than OTM calls with the same distance k from strike price
- Meaning the Implied Volatility for Puts are higher than calls

*see long risk reversal strategy at Section 2.12 (directional strategy)*

### Section 7.6 of "151 Trading Strategies" -- Variance Swaps

- Variance Swaps is a less cumbersome & costly alternative to Delta Hedging
- Variance Swaps -> OTC forward contract on future realized variance
- P(T) = N × (v(T) − K), where,
- P(T) = your payoff
- N = amount you're trading
- v(T) = realized variance (or volatility)
- K = variance strike (agreed ahead of time)

## Documentation/Code - Available Upon Request
<!--
* Presentation: https://drive.google.com/file/d/1R2Fk2iDnc08wDnJ_krb8MJOiw7oCG0I8/view?usp=drive_link
* Research Paper: https://drive.google.com/file/d/1pedYSsV9Ro9v1JTKnI4mTHzlndzENcLc/view?usp=drive_link
-->