Bayesian Structured Time Series Analysis with Parallel Tempering for Stock Market Prediction

Overview

An experimental project which implemented the Bayesian structured time series (BSTS) model using Langevin-gradients parallel tempering.
Markov chain Monte Carlo (MCMC) methods were implemented in a parallel computing environment.
Compare the stock price forecasting model with state-of-art neural network training algorithms (FNN-SGD and FNN-Adam)

Specifications

data.py - Used for data preprocessing.\
ann.py - Desired parameters should be set in the artificial neural network to run the results.

Output

Following are some sample results of MMM’s stock price prediction.
These are one-step, two-step, five-step prediction result and error analysis respectively.
The grey area is the uncertainty of the prediction results.

Figure 1: Sample outputs of MMM's stock price prediction 70 days from the time of analysis.