Use unsupervised and supervised learning to predict stocks

Overview

AIAlpha: Multilayer neural network architecture for stock return prediction

forthebadge made-with-python

GitHub license PRs Welcome

This project is meant to be an advanced implementation of stacked neural networks to predict the return of stocks. My goal for the viewer is to understand the core principles that go behind the development of such a multilayer model and the nuances of training the individual components for optimal predictive ability. Once the core principles are understood, the various components of the model can be replaced with the state of the art models available at time of usage.

The workflow is similar to the approach in the excellent text Advances in Financial Machine Learning by Marcos Lopez de Prado, which I recommend to anyone who wants to learn about applying machine learning techniques to financial data. The data that was used for this project is not in the repository due to size constraints in GitHub, but the raw data was open sourced from Tick Data LLC, but now I believe is not available.

In essense, we will be making bars (tick, volume or dollar) based on the tick data, apply feature engineering, reduce the dimensions using an autoencoder and finally use a machine learing model to make predictions. I have implemented both a LSTM regression model and a Random Forest classification model to classify the direction of the move.

This model is not meant to be used to live trade without modifications. However, an extended version of this model can very well be profitable with the right strategies.

I truly hope you find this project informative and useful in developing your own trading strategies or machine learning models.

This project illustrates how to use machine learning to predict the future prices of stocks. In order to efficiently allocate the capital to those stocks, check out OptimalPortfolio

Disclaimer, this is purely an educational project. Any backtesting performance do not guarentee live trading results. Trade at your own risk. This is only a guide on the usage of the model. If you want to delve into the reasoning behind the model and the theory, please check out my blog: Engineer Quant

Contents

Overview

Those who have done some form of machine learning would know that the workflow follows this format: acquire data, preprocess, train, test, monitor model. However, given the complexity of this task, the workflow has been modified to the following:

  1. Acquire the tick data - this is the primary data for our model.
  2. Preprocess the data - we need to sample the data using some method. Subsequently, we make the train-test splits.
  3. Train the stacked autoencoder - this will give us our feature extractor.
  4. Process the data - this will give us the features of our model, along with train, test datasets.
  5. Use the neural network/random forest to learn from the training data.
  6. Test the model with the testing set - this gives us a gauge of how good our model is.

Now let me elaborate the various parts of the pipeline.

Quickstart

For those who just want to see the model work, run the following code (make sure you are on Python 3 to prevent any bugs or errors):

pip install -r requirements.txt
python run.py

Note: Due to GitHub file size restrictions, I have only uploaded part of the data (1 million rows), so the model results may vary from the one shown below.

Bar Sampling

Running machine learning algorithms, or any other statistical models, directly on tick level data often leads to poor results, due to the high level of noise caused by the bid-ask bounce, and the high nonlinearity in the nature of the data. Therefore, we need to sample the data at some interval (which can be decided depending on the frequency of the predictive model). The sampling that we are used to seeing is time sampled (we get bars every 1min), but this is known to exhibit non stationarities and the returns are not normally distributed. So, as explained in Advances in Financial Machine Learning, we are going to sample it according to the number of ticks, or the amount of volume or the amount of dollars traded. These bars show better statistical properties and are preferable for machine learning applications.

Feature Engineering

Given our OHLCV data from our sampling procedure, we can go ahead and create features that we feel might add information to the forecast. I have constructed a set of features that are based on moving averages and rolling volatilities of the various prices and volumes. This set of features can be extended accordingly.

Stacked Autoencoder

Given our features, we notice that the dimension of the dataset is huge (185 for my configuration). This can pose a lot of problems when we run machine learning algorithms due to the curse of dimensionality. However, we can attempt to overcome this by using neural networks that are able to decompress the data given into smaller number of neurons than the input number. When we train such a neural network, it becomes able to extract the 'important sections' of the data so to speak. Hence, this compressed version of the data can be considered as features. Although this method is useful, the downside is that we do not know what the various compressed data points mean and hence cannot extract methods to achieve them in differnt datasets.

Neural Network Model

Using neural networks for the prediction of time series has become widespread and the power of neural networks is well known. I have used a LSTM model for its memory property. However, an issue I faced with the training of the neural network model is that there was a tendency for the model to fit to a constant, as it turned out to be a local minima for the loss function. One way to overcome this is using different initialisations for the weights, and tuning the hyperparameters.

Random Forest Model

Sometimes, it might be better to use a simpler model as apposed to a sophisticated neural network. This is especially true when the amount of data available is not enough for deep models. Even though I used tick level data, the dataset was only around 5 million rows. After sampling, the number of rows drops and it is not enough for deep learning models to learn effectively from. So, I wanted to use a random forest classification model that classified the direction of the next bar.

Results

Using this stacked neural network model, I was able to achieve decent results. The following are graphs of my predictions vs the actual market prices for various securities.

EURUSD

alt text

EURUSD prices - R^2: 0.90

alt text

For the random forest classification model, the results were better. I used tick bars for this simulation.

The base case used is merely predicting no moves in the market. The out of sample results were:

Tick bars:
    Model log loss: 2.78
    Base log loss: 4.81

Volume bars:
    Model log loss: 1.69
    Base log loss: 5.06

Dollar bars:
    Model log loss: 2.56
    Base log loss: 2.94

It is also useful to understand how much of an impact the autoencoders made, so I ran the model without autoencoders and the results were:

Tick bars:
    Model log loss: 5.12
    Base log loss: 4.81

Volume bars:
    Model log loss: 3.25
    Base log loss: 5.06

Dollar bars:
    Model log loss: 3.62
    Base log loss: 2.94

Online Learning

The training normally stops after the model has trained on historic data and merely predicts future data. However, I believe that it might be a waste of data if the model does not also learn from the predictions. This is done by training the model on the new (prediction, actual) pairs to continually improve the model.

What's next?

The beauty of this model is the once the construction is understood, the individual models can be swapped out for the best model there is. So over time the actual models used here will be different but the core framework will still be the same. I am also working on improving the current model with ideas from Advanced in Financial Machine Learning, such as adding sample weights, cross-validation and ensemble techniques.

Contributing

I am always grateful for feedback and modifications that would help!

Hope you have enjoyed that! To see more content like this, please visit: Engineer Quant

Owner
Vivek Palaniappan
Keen on finding effective solutions to complex problems - looking into the broad intersection between engineering, finance and AI.
Vivek Palaniappan
Learning Optical Flow from a Few Matches (CVPR 2021)

Learning Optical Flow from a Few Matches This repository contains the source code for our paper: Learning Optical Flow from a Few Matches CVPR 2021 Sh

Shihao Jiang (Zac) 159 Dec 16, 2022
Object detection and instance segmentation toolkit based on PaddlePaddle.

Object detection and instance segmentation toolkit based on PaddlePaddle.

9.3k Jan 02, 2023
ELECTRA: Pre-training Text Encoders as Discriminators Rather Than Generators

ELECTRA Introduction ELECTRA is a method for self-supervised language representation learning. It can be used to pre-train transformer networks using

Google Research 2.1k Dec 28, 2022
An unofficial styleguide and best practices summary for PyTorch

A PyTorch Tools, best practices & Styleguide This is not an official style guide for PyTorch. This document summarizes best practices from more than a

IgorSusmelj 1.5k Jan 05, 2023
Official PyTorch Implementation of paper "NeLF: Neural Light-transport Field for Single Portrait View Synthesis and Relighting", EGSR 2021.

NeLF: Neural Light-transport Field for Single Portrait View Synthesis and Relighting Official PyTorch Implementation of paper "NeLF: Neural Light-tran

Ken Lin 38 Dec 26, 2022
CenterFace(size of 7.3MB) is a practical anchor-free face detection and alignment method for edge devices.

CenterFace Introduce CenterFace(size of 7.3MB) is a practical anchor-free face detection and alignment method for edge devices. Recent Update 2019.09.

StarClouds 1.2k Dec 21, 2022
Bling's Object detection tool

BriVL for Building Applications This repo is used for illustrating how to build applications by using BriVL model. This repo is re-implemented from fo

chuhaojin 47 Nov 01, 2022
New approach to benchmark VQA models

VQA Benchmarking This repository contains the web application & the python interface to evaluate VQA models. Documentation Please see the documentatio

4 Jul 25, 2022
Code for Mining the Benefits of Two-stage and One-stage HOI Detection

Status: Archive (code is provided as-is, no updates expected) PPO-EWMA [Paper] This is code for training agents using PPO-EWMA and PPG-EWMA, introduce

OpenAI 33 Dec 15, 2022
Provide partial dates and retain the date precision through processing

Prefix date parser This is a helper class to parse dates with varied degrees of precision. For example, a data source might state a date as 2001, 2001

Friedrich Lindenberg 13 Dec 14, 2022
Code release of paper "Deep Multi-View Stereo gone wild"

Deep MVS gone wild Pytorch implementation of "Deep MVS gone wild" (Paper | website) This repository provides the code to reproduce the experiments of

François Darmon 53 Dec 24, 2022
Code for Phase diagram of Stochastic Gradient Descent in high-dimensional two-layer neural networks

Phase diagram of Stochastic Gradient Descent in high-dimensional two-layer neural networks Under construction. Description Code for Phase diagram of S

Rodrigo Veiga 3 Nov 24, 2022
Datasets, tools, and benchmarks for representation learning of code.

The CodeSearchNet challenge has been concluded We would like to thank all participants for their submissions and we hope that this challenge provided

GitHub 1.8k Dec 25, 2022
Optical machine for senses sensing using speckle and deep learning

# Senses-speckle [Remote Photonic Detection of Human Senses Using Secondary Speckle Patterns](https://doi.org/10.21203/rs.3.rs-724587/v1) paper Python

Zeev Kalyuzhner 0 Sep 26, 2021
3rd place solution for the Weather4cast 2021 Stage 1 Challenge

weather4cast2021_Stage1 3rd place solution for the Weather4cast 2021 Stage 1 Challenge Dependencies The code can be executed from a fresh environment

5 Aug 14, 2022
Neighborhood Reconstructing Autoencoders

Neighborhood Reconstructing Autoencoders The official repository for Neighborhood Reconstructing Autoencoders (Lee, Kwon, and Park, NeurIPS 2021). T

Yonghyeon Lee 24 Dec 14, 2022
Instance Segmentation by Jointly Optimizing Spatial Embeddings and Clustering Bandwidth

Instance segmentation by jointly optimizing spatial embeddings and clustering bandwidth This codebase implements the loss function described in: Insta

209 Dec 07, 2022
Regression Metrics Calculation Made easy for tensorflow2 and scikit-learn

Regression Metrics Installation To install the package from the PyPi repository you can execute the following command: pip install regressionmetrics I

Ashish Patel 11 Dec 16, 2022
PyTorch Implementation of SSTNs for hyperspectral image classifications from the IEEE T-GRS paper "Spectral-Spatial Transformer Network for Hyperspectral Image Classification: A FAS Framework."

PyTorch Implementation of SSTN for Hyperspectral Image Classification Paper links: SSTN published on IEEE T-GRS. Also, you can directly find the imple

Zilong Zhong 54 Dec 19, 2022
Level Based Customer Segmentation

level_based_customer_segmentation Level Based Customer Segmentation Persona Veri Seti kullanılarak müşteri segmentasyonu yapılmıştır. KOLONLAR : PRICE

Buse Yıldırım 6 Dec 21, 2021