MSc project: Use Large Artificial Neural Nets to Extract Semantic Signals for Long-term Sequential Forecasting

multi-step (30-days) forecast the price of sterling in Chinese yuan

repository on GitHub

creater: Yishan Liu

introduction

• this is a collection of the materials I used for my MSc proejct on long term financial time series forecasting with deep learning.
• this collection includes the data I used, my paper, and my python implementaiton.
• I am looking forward for your feedbacks, thank you!

compare 5 RNN encoder-decoders and their non encoder-decoder versions

• the 5 RNN encoder-decoders are:
  1. GRU encoder-decoder
  2. bidirectional GRU encoder-decoder
  3. CNN-GRU encoder-decoder
  4. convLSTM encoder-decoder
  5. bidirectional convLSTM encoder-decoder

$${\color{red} I \space created \space a \space new \space architecture: \space bidirectional \space convLSTM \space encoder-decoder}$$

• the 5 non encoder-decoders are:
  1. GRU
  2. bidirectional GRU
  3. CNN-GRU
  4. convLSTM
  5. bidirectoinal convLSTM

visualised result from the best performing model:

• evaluation metrics:

  • train mse: 0.0361340157687664
  • test mse: 0.017722204327583313
  • train rmse: 0.17656706273555756
  • test rmse: 0.12243817746639252

• other statistics - means:

  • training data mean: 8.841544827586207
  • train prediction mean: 8.791829
  • testing data mean: 8.717213409961685
  • test prediction mean: 8.674898

• other statistics - standard deviations:

  • training data standard deviation: 0.17639548961572413
  • train prediction standard deviation: 0.0813758
  • testing data standard deviation: 0.14175250034034664
  • test prediction standard deviation: 0.06592484