kevindirect / thesis-code

An Empirical Evaluation of Neural Process Meta-Learners for Financial Forecasting

Thesis Code

Thesis Details

• Title: An Empirical Evaluation of Neural Process Meta-Learners for Financial Forecasting
• Author: Kevin Patel

Running

• dump basic preprocessed data: data/Preproc/run.sh
• ANP volatiltiy model nowcast hyperparam tuning over train/val: run-exp-000.bash
• ANP volatiltiy model forecast hyperparam tuning over train/val: run-exp-001.bash
• ANP volatiltiy model forecast (fixed hyperparams) over train/val/test: run-exp-002.bash
• simulated trading model over train/val: model/TradingModel/run.sh
• simulated trading model over train/val/test: model/TradingModel/run-final.sh
• basic smoketests: smoke-*.sh

Overview

• This branch, master/data-frd-minutely, has all the code submitted for the thesis
• The branch data-tr-hourly is for a previous iteration of the project
• Unfortunately raw data used to generate the results was purchased from FirstRate Data LLC, so it cannot be shared

data/

• has all the raw and preprocessed data
• preprocessing is done from Julia scripts/Pluto.jl notebooks in the Preproc package
• preprocessed data can be loaded in Python via a PytorchLightning DataModule

model/

• has all the model code (Pytorch models wrapped in PytorchLightning)

RV Model

• {model, np}_util.py contain Pytorch model classes
• pl_{generic, np}.py are LightningModule classes that wrap Pytorch models for PytorchLightning
• expo.py is the optuna hyperparameter optimizing runner
• expm.py is the manual/fixed hyperparameter experiment runner
• the model/exp-<proc>-<data> directories contain completed realized volatility trial results
• hyperparameter sets are stored on disk in json files

Trading Model

• model/TradingModel contains the julia package with the trading model code
• Trading model uses features/predictions dumped from the last realized volatiltiy models
• model/tm-<proc>-<data> directories contain completed trading model results

common_util.py

• common_util.py contains functions, data structures, and classes used throughout the project
• Much of the code in here is not used anymore (previous iterations of project)

Other

• Most code is arranged in python subpackages, <subpackage>/common.py contains common constants, defaults, and utilities
• Subpackage scripts are run by running them as modules (using the -m flag), see the shell scripts at the project root for examples