LiquidityShock

We implement a Stanford study on determining the best bid/ask following a liquidity shock (a large trade sweeping through several levels of the orderbook). We address the original paper's high-bias problem by designing and implementing a non-linear transformer in PyTorch to forecast whether or not the best bid/ask reverts to its original price or is significantly altered. We first follow the paper's methodology of clustering the training data using a K-means algorithm and then training a Gaussian-Discriminant Analysis and Linear Regression to benchmark our results. We then design and implement a Sequence-to-Sequence regression transformer for our main forecasting model to improve on the paper's methods.