1 of 10

Optiver - Trading at the Close

By: Matthew Durai, Reetom Gangopadhyay, Ryan Gilbert

Quant ML Competition

Premier academic organization for aspiring finance professionals.

2 of 10

Competition Overview

  • Develop a model capable of predicting stock closing price movements using data from the order book and the closing auction of stocks.
  • Information from the auction can be used to adjust prices, assess supply and demand dynamics, and identify trading opportunities.

Optiver - Trading at the Close

2

Premier academic organization for aspiring finance professionals.

3 of 10

Competition Overview cont.

  • Active training phase and a second period where models will be run against new market data
  • 5.24 million rows of orderbook info for the closing 10 minutes of NASDAQ stocks, predict future move relative to synthetic index
  • Target move in basis points (0.01%), Leaderboard MAE
  • ~4,000 teams

Optiver - Trading at the Close

3

Premier academic organization for aspiring finance professionals.

4 of 10

Data

Optiver - Trading at the Close

4

Columns:

  • Imbalance size
  • Reference price
  • Match size
  • Bid/Ask prices
  • Imbalance direction
  • Seconds in bucket
  • Far/Near price
  • Bid/Ask sizes

Premier academic organization for aspiring finance professionals.

5 of 10

Feature Engineering

  • Combinations of order book information
  • Windowed features
  • Rolling averages, standard deviations
  • Mean, standard deviation, skew, kurtosis for price features
  • Time related features

Optiver - Trading at the Close

5

Premier academic organization for aspiring finance professionals.

6 of 10

LightGBM

  • LightGBM is a framework developed to implement a Gradient Boosting Machine.
    • Advantages: lightweight, fast, supports training distribution, built-in features to reduce overfitting.
    • Disadvantages: Highly sensitive to hyperparameter tuning, difficult to interpret results, highish memory usage.

​

In general, LightGBM had the strongest performance out of all of the models we tried such as XGBoost and CatBoost.

Optiver - Trading at the Close

6

Premier academic organization for aspiring finance professionals.

7 of 10

CV Strategy and Ensemble Approach

CV strategy: Purged 5-Fold

  • Find optimal hyperparameters
  • Early stopping to prevent overfit
  • Performance based on MAE

​

Ensemble approach:

  1. Rerun Purged 5-Fold
  2. Run model on complete dataset
  3. Combine forecasts of all 6 models

​

Weights on forecasts based on performance on validation set.

​

Bates-Granger Equation:

​

​

� Modified Equation:

�

Optiver - Trading at the Close

​

, where is 1/MSE

, where is 1/MAE2

7

Premier academic organization for aspiring finance professionals.

8 of 10

Results & Limitations

MAE: 5.3364, Rankings: 187/4034

Limitations:

  1. Optimization difficulties
  2. Runtime limits (9 hr)
  3. Hardware limits (RAM)
  4. Kaggle GPU limits (30hr/week)
  5. API limitations (lags of target not available)

​

Optiver - Trading at the Close

​

8

Premier academic organization for aspiring finance professionals.

9 of 10

Reflections & What’s Next

​

  • Real world problem, competing against professionals
  • Memory optimization
  • Working with bigger datasets

​

  • Neural network, RNN, …
  • Hyperparameter tuning
  • Feature engineering with importances

Optiver - Trading at the Close

​

9

Premier academic organization for aspiring finance professionals.

10 of 10

Optiver - Trading at the Close

10

Q+A

Premier academic organization for aspiring finance professionals.