search

huseinzol05 / Stock-Prediction-Models

Gathers machine learning and deep learning models for Stock forecasting including trading bots and simulations
Apache License 2.0
8.09k stars 2.84k forks source link
deep-learning deep-learning-stock evolution-strategies learning-agents lstm lstm-sequence monte-carlo monte-carlo-markov-chain seq2seq stock-market stock-prediction-models stock-price-forecasting stock-price-prediction strategy-agent trading-bot

readme

logo

MIT License

Stock-Prediction-Models, Gathers machine learning and deep learning models for Stock forecasting, included trading bots and simulations.

Table of contents

Contents

Models

Deep-learning models

  1. LSTM
  2. LSTM Bidirectional
  3. LSTM 2-Path
  4. GRU
  5. GRU Bidirectional
  6. GRU 2-Path
  7. Vanilla
  8. Vanilla Bidirectional
  9. Vanilla 2-Path
  10. LSTM Seq2seq
  11. LSTM Bidirectional Seq2seq
  12. LSTM Seq2seq VAE
  13. GRU Seq2seq
  14. GRU Bidirectional Seq2seq
  15. GRU Seq2seq VAE
  16. Attention-is-all-you-Need
  17. CNN-Seq2seq
  18. Dilated-CNN-Seq2seq

Bonus

  1. How to use one of the model to forecast t + N, how-to-forecast.ipynb
  2. Consensus, how to use sentiment data to forecast t + N, sentiment-consensus.ipynb

Stacking models

  1. Deep Feed-forward Auto-Encoder Neural Network to reduce dimension + Deep Recurrent Neural Network + ARIMA + Extreme Boosting Gradient Regressor
  2. Adaboost + Bagging + Extra Trees + Gradient Boosting + Random Forest + XGB

Agents

  1. Turtle-trading agent
  2. Moving-average agent
  3. Signal rolling agent
  4. Policy-gradient agent
  5. Q-learning agent
  6. Evolution-strategy agent
  7. Double Q-learning agent
  8. Recurrent Q-learning agent
  9. Double Recurrent Q-learning agent
  10. Duel Q-learning agent
  11. Double Duel Q-learning agent
  12. Duel Recurrent Q-learning agent
  13. Double Duel Recurrent Q-learning agent
  14. Actor-critic agent
  15. Actor-critic Duel agent
  16. Actor-critic Recurrent agent
  17. Actor-critic Duel Recurrent agent
  18. Curiosity Q-learning agent
  19. Recurrent Curiosity Q-learning agent
  20. Duel Curiosity Q-learning agent
  21. Neuro-evolution agent
  22. Neuro-evolution with Novelty search agent
  23. ABCD strategy agent

Data Explorations

  1. stock market study on TESLA stock, tesla-study.ipynb
  2. Outliers study using K-means, SVM, and Gaussian on TESLA stock, outliers.ipynb
  3. Overbought-Oversold study on TESLA stock, overbought-oversold.ipynb
  4. Which stock you need to buy? which-stock.ipynb

Simulations

  1. Simple Monte Carlo, monte-carlo-drift.ipynb
  2. Dynamic volatility Monte Carlo, monte-carlo-dynamic-volatility.ipynb
  3. Drift Monte Carlo, monte-carlo-drift.ipynb
  4. Multivariate Drift Monte Carlo BTC/USDT with Bitcurate sentiment, multivariate-drift-monte-carlo.ipynb
  5. Portfolio optimization, portfolio-optimization.ipynb, inspired from https://pythonforfinance.net/2017/01/21/investment-portfolio-optimisation-with-python/

Tensorflow-js

I code LSTM Recurrent Neural Network and Simple signal rolling agent inside Tensorflow JS, you can try it here, huseinhouse.com/stock-forecasting-js, you can download any historical CSV and upload dynamically.

Misc

  1. fashion trending prediction with cross-validation, fashion-forecasting.ipynb
  2. Bitcoin analysis with LSTM prediction, bitcoin-analysis-lstm.ipynb
  3. Kijang Emas Bank Negara, kijang-emas-bank-negara.ipynb

Results

Results Agent

This agent only able to buy or sell 1 unit per transaction.

  1. Turtle-trading agent, turtle-agent.ipynb
  1. Moving-average agent, moving-average-agent.ipynb
  1. Signal rolling agent, signal-rolling-agent.ipynb
  1. Policy-gradient agent, policy-gradient-agent.ipynb
  1. Q-learning agent, q-learning-agent.ipynb
  1. Evolution-strategy agent, evolution-strategy-agent.ipynb
  1. Double Q-learning agent, double-q-learning-agent.ipynb
  1. Recurrent Q-learning agent, recurrent-q-learning-agent.ipynb
  1. Double Recurrent Q-learning agent, double-recurrent-q-learning-agent.ipynb
  1. Duel Q-learning agent, duel-q-learning-agent.ipynb
  1. Double Duel Q-learning agent, double-duel-q-learning-agent.ipynb
  1. Duel Recurrent Q-learning agent, duel-recurrent-q-learning-agent.ipynb
  1. Double Duel Recurrent Q-learning agent, double-duel-recurrent-q-learning-agent.ipynb
  1. Actor-critic agent, actor-critic-agent.ipynb
  1. Actor-critic Duel agent, actor-critic-duel-agent.ipynb
  1. Actor-critic Recurrent agent, actor-critic-recurrent-agent.ipynb
  1. Actor-critic Duel Recurrent agent, actor-critic-duel-recurrent-agent.ipynb
  1. Curiosity Q-learning agent, curiosity-q-learning-agent.ipynb
  1. Recurrent Curiosity Q-learning agent, recurrent-curiosity-q-learning.ipynb
  1. Duel Curiosity Q-learning agent, duel-curiosity-q-learning-agent.ipynb
  1. Neuro-evolution agent, neuro-evolution.ipynb
  1. Neuro-evolution with Novelty search agent, neuro-evolution-novelty-search.ipynb
  1. ABCD strategy agent, abcd-strategy.ipynb

Results signal prediction

I will cut the dataset to train and test datasets,

  1. Train dataset derived from starting timestamp until last 30 days
  2. Test dataset derived from last 30 days until end of the dataset

So we will let the model do forecasting based on last 30 days, and we will going to repeat the experiment for 10 times. You can increase it locally if you want, and tuning parameters will help you by a lot.

  1. LSTM, accuracy 95.693%, time taken for 1 epoch 01:09
  1. LSTM Bidirectional, accuracy 93.8%, time taken for 1 epoch 01:40
  1. LSTM 2-Path, accuracy 94.63%, time taken for 1 epoch 01:39
  1. GRU, accuracy 94.63%, time taken for 1 epoch 02:10
  1. GRU Bidirectional, accuracy 92.5673%, time taken for 1 epoch 01:40
  1. GRU 2-Path, accuracy 93.2117%, time taken for 1 epoch 01:39
  1. Vanilla, accuracy 91.4686%, time taken for 1 epoch 00:52
  1. Vanilla Bidirectional, accuracy 88.9927%, time taken for 1 epoch 01:06
  1. Vanilla 2-Path, accuracy 91.5406%, time taken for 1 epoch 01:08
  1. LSTM Seq2seq, accuracy 94.9817%, time taken for 1 epoch 01:36
  1. LSTM Bidirectional Seq2seq, accuracy 94.517%, time taken for 1 epoch 02:30
  1. LSTM Seq2seq VAE, accuracy 95.4190%, time taken for 1 epoch 01:48
  1. GRU Seq2seq, accuracy 90.8854%, time taken for 1 epoch 01:34
  1. GRU Bidirectional Seq2seq, accuracy 67.9915%, time taken for 1 epoch 02:30
  1. GRU Seq2seq VAE, accuracy 89.1321%, time taken for 1 epoch 01:48
  1. Attention-is-all-you-Need, accuracy 94.2482%, time taken for 1 epoch 01:41
  1. CNN-Seq2seq, accuracy 90.74%, time taken for 1 epoch 00:43
  1. Dilated-CNN-Seq2seq, accuracy 95.86%, time taken for 1 epoch 00:14

Bonus

  1. How to forecast,
  1. Sentiment consensus,

Results analysis

  1. Outliers study using K-means, SVM, and Gaussian on TESLA stock
  1. Overbought-Oversold study on TESLA stock
  1. Which stock you need to buy?

Results simulation

  1. Simple Monte Carlo
  1. Dynamic volatity Monte Carlo
  1. Drift Monte Carlo
  1. Multivariate Drift Monte Carlo BTC/USDT with Bitcurate sentiment
  1. Portfolio optimization