(paper) Transfer Learning for Financial Time Series Forecasting

Transfer Learning for Financial Time Series Forecasting

Contents

1. Abstract
2. TL in TS Forecasting
3. Method
   1. Training Strategy
   2. Network Architecture

0. Abstract

poor performance of applying DL models in short TS

\(\rightarrow\) solve with Transfer Learning

( + propose similiarity-based approach for selecting source datasets )

Setting

• 2 source datasets
• 1 target datsets

1. TL in TS Forecasting

Recent works

• Ye et al [27] : propose a novel transfer learning for TS forecasting

• Fawaz et al [8] : calculate similarity to select source datasets

  • propose DTW based algorithm

• Hu et al [14] : pre-train a 2-layer DNN

  • parameters of model are shared with all TS

    ( = universal feature transformation )

Proposed work

• 1st layer DNN : train with one source dartaset
• 2nd layer DNN : train with both source datasets

\(\rightarrow\) not only universal features, but also specific features

2. Method

(1) Training Strategy

• multi-domain TL
• all source domains *may not have same influence on target domain

This paper :

• not only contain universal features
• but also maintain specific features of the source domains

(2) Network Architecture

1. 2-layer DNN
2. 2-layer LSTM