vishwesh4 / bdc_shaastra

Used LSTM in Big Data Challenge competiton hosted by Shaastra IIT Madras. Won the second prize

Big Data Challenge Shaastra

Big data challenge by HSBC hosted by Shaastra. The problem statement being, to predict price after 30 seconds in a limit order book.

Getting Started

These instructions will get you a copy of the project up and running on your local machine for development and testing purposes.

Prerequisites

Libraries needed

re
time
datetime
operator
numpy
pandas
collections
unicodedata
collections
seaborn
matplotlib
tqdm
IPython.display.Image
statsmodels
sklearn
keras
Tensorflow
scipy

Approach

In general, a set of fully connected layers are attached and used to classify the input data. However, all inputs to the fully connected layers are assumed independent of each other, while our features still contain time dependencies. In order to capture temporal relationship that exist in the extracted features, we replace the fully connected layers with LSTM units. The activation of a LSTM unit is fed back to itself and the memory of past activations is kept with a separate set of weights, so the temporal dynamics of our features can be modelled. We use 3 layers LSTM in our work, and find using more units can lead to overfitting. In our model we predict the stock price 30 seconds ahead by actually finding the price movement happening after 30 secs. We labelled it into 3 classes.

1. +1 representing tick up
2. 0 representing no significant change
3. -1 representing tick down The last output layer uses a softmax activation function to classify into 3 classes and hence the final output elements represent the probability of each price movement class (tick up, tick down or no change) at each time step.

For prediction of predMid, we used:

Labelling

Because financial data is highly stochastic, if we simply compare pt and pt+k to decide the price movement, the resulting label set will be noisy. We adopt the idea of introducing a smoothed labelling method. First, m− denotes the mean of the previous k mid-prices and m+ denotes the mean of the next k mid-prices:

where pt is the mid-price defined in Equation (1) and k is the prediction horizon (k = 300 signifying 30 seconds on average). Then, we use m− and m+ to define the direction of price movement (lt) at time t by:

where the threshold α determines the smallest change in price that must occur for the mid-price to be considered upward (+1) or downward (−1). We choose the threshold α such that all the classes are balanced for each instrument during the training period. This can be understood as adjusting for the average volatility of a given instrument, i.e. if a stock is more volatile we require a larger upwards (downwards) price move to classify it as a +1 (−1).

Feature Description

Running the tests

You can load data from :-

• 1 - intraday/quote_in.csv - data for quotes, training set
• 2 - intraday/quote_out.csv - data for quotes, test set
• 3 - intraday/trade_in.csv - trade data , training set
• 4 - intraday/trade_out.csv - trade data , test set

We have some important files

• 1 - model.ipynb - Main pipeline
• 2 - core/model.py - contains class model which is our LSTM model
• 3 - core/data_processor.py - data_loader class for ease of loading big data
• 4 - config.json - contains all the important parameters for model
• 5 - saved_models - contains the saved checkpoints of both prediction and model weights
• 6 - model.h5 - contains the current model weights we are using for replicability