cfm_data_challenge_ens

Where will the next trade take place? : https://challengedata.ens.fr/participants/challenges/40/

Context of the challenge

Most financial markets use an electronic trading mechanism called a limit order book to facilitate the trading of assets (stocks, futures, options, etc.). Participants submit (or cancel) orders to this electronic order book. These orders are requests to buy or sell a given quantity of an asset at a specified price, thus allowing buyers to be matched with sellers at a mutually agreed price. Since an asset can be traded on multiple trading venues, participants can choose to which venue they send an order. For instance, a US stock can be traded on various exchanges, such as NYSE, NASDAQ, Direct Edge or BATS. When sending an order, participants generally select the best available trading venue at that time. Their decisions may include a statistical analysis of past venue activity.

Given recent trades and order books from a set of trading venues, predict on which trading venue the next trade will be executed.

My Approach

The approach I took in this challenge is rather classic. It consists in a in-depth feature engineering followed by a well-tuned XGBoost trained using GPUs.

Preprocessing

Not a lot of preprocessing was needed since the data was rather clean.

• I removed the stocks that were not in the test set
• I renamed columns from tuple to string to ease manipulation
• There were a few nans in the OB price. It means that no price is proposed at that point in time. To remove them, I replaced the price with a price very far from the mid. This was especially useful to make new features and avoid "propagating" nans
• Normalizing trade time w.r.t. the last trade in the list, so that trade times are comparable across samples
• Normalize last update of the orders books, to make them comparable

Feature engineering

Feature engineering was (I think) the main part of the challenge. In order to predict where the next trade would go, I used a series of roughly 260 different features :

Features on last 10 trades

• Count the frequency of each venue among the last 3, 5 and 10 trades. This is quite an intuitive feature: if a venue was much used before, it will probably continue after.
• Count the frequency of each venue among the trades that happened in the last 0.1, 1 and 10 seconds. This allow to capture only near trades, which makes them more relevant.
• Descriptive features on those trades: mean, std, min max of price and quantity, number of trades on ask, std of the times of the trades (not very useful)
• Stats on the biggest and best price trade among 10 last (not very useful)

Features on order book size, price and update

• Compute the total book size for bid and ask and for both levels
• Rank OB by proposed price for both levels of the book (one of the most powerful features). It allows to create a "stationary" feature that does not depend on the context/time of the day, to describe how attractive this order book is w.r.t others.
• Prices of both level versus best price available on 6 venues (some kind of normalization/ratio that allow features to be comparable)
• Some weighted quantity-price index
• Ratio between the size / price of the first and second level to measure the gap between them

Target/Frequency encoding (may have caused severe problems on the private set)

• Encode each stock / day / time of the day by the frequence of the trades on each venue on the train data (ie groubpy stock or day or time of the day and compute the frequency for each venue) and use it as a feature. It adds 6 columns per feature encoded.

Training

The training procedure was rather standard

• Cross-validation : 5-Folds (using sklearn implementation)
• Memory reduction trick to allow all ~260 features to fit in 16Go of RAM
• Model : XGBoostRegressor (from xgboost package) with following parameters:
  • using GPU (~45 min)
  • 0.05 learning rate
  • 8 max depth
  • 0.6 columns subsampling (nice boost)
  • 4 min_child_weight

Final Score

Validation score was 0.5130, which translated in 0.5040 on the public LB (4th) and, for some strange reason, in 0.2610 on the public LB. This difference is quite difficult to understand, because even by overfitting "voluntarily" I couldn't get such a modification (and 0.2610 is way behind the naive baseline that consists in predicting always the last venue). Moreover, the public-private score was quite stable for all other candidates

Description of the data

For each row, we want predict on which venue the next trade will be executed. The stock is represented by a randomized stock_id and the day by a randomized day_id. Each row provides a description of 6 order books, from 6 trading venues, and a history of 10 trades for the corresponding asset

Order books

An order book lists the quantities of an asset that are currently on offer by sellers (who ask for higher prices) and the quantities that buyers wish to acquire (who bid at lower prices). The six order books (one for each trading venue) are described in the dataset through the best two bids and best two asks (which makes them respectively the two highest bid prices of the buyers and the two lowest ask prices of the sellers)

Each of the six books is described as follows:

• The bid column (resp. 'ask') represents the difference between the best bid (resp. best ask) and the aggregate mid-price, expressed in some fixed currency unit. The bid1 column (resp. 'ask1') represents the difference between the second best bid (resp. second best ask) and the aggregate mid-price,

• The bid_size column (resp. 'ask_size') represents the total number of stocks available at the best bid (resp. best ask) divided by the aggregate volume. The bid_size1 column (resp. 'ask_size1') represents the total number of stocks available at the second best bid (resp. at the second best ask) divided by the aggregate volume,

• The 'ts_last_update' column corresponds to the timestamp of the last update of the book.

Trades

Each row also comprises a description of the ten last trades for the corresponding asset. A trade represents a transaction of a certain quantity of an asset at a given price between a buyer and a seller. The ten trades from the history of trades are given are described as follows:

• Its quantity (qty): the number of stocks traded, divided by the aggregate volume,
• Its timestamp (tod): when the trade was executed,
• Its price (price), representing the difference between the trade price with the aggregate mid-price,
• Its source (source_id) representing the trading venue on which this particular trade was executed.