Skip to content
All library documents

Feature Selection for Deep Neural Networks in Financial Prediction

Article Quant Q&A · Author: guyov
Read the original

Summary

The document considers feature choice for a multilayer perceptron that classifies financial price moves as buy, hold, or sell. The proposed inputs concatenate recent returns and transaction volumes, and the question is whether a vector containing more than fifty elements explains poor predictive results.

The responses offer two limited pointers: input size alone is not enough to diagnose the problem, and feature selection can be incorporated into a deep network using a deep-lasso approach. The cited comparison to a much larger input vector is illustrative, not evidence that a larger feature set will work well for financial prediction. No experiments, validation results, or practical selection procedure are provided, so the exchange does not establish whether the model’s poor performance comes from features, architecture, training, or another cause.

Key ideas

  • The example classifier uses recent returns and transaction volume as inputs to buy, hold, and sell outputs.
  • Input dimension alone does not establish whether a model’s feature set is suitable.
  • Deep feature selection methods such as deep lasso can select inputs within a neural network.
  • The discussion provides suggestions but no financial prediction results or diagnostic procedure.

Tags

Machine learningEquitiesStatistics
Full text
# Feature Selection Effect on Deep Multi-Layer-Perceptron for Financial Applications


# Feature Selection Effect on Deep Multi-Layer-Perceptron for Financial Applications












I am trying to build a machine learning system for financial price prediction. I am using a 3 layer MLP (a deep network) with 3 outputs (buy,hold,sell).

I am using different features such as price and volume. In order to take into account past behavior, I concatenate the price and volume data into one vector. The first part of the vector is the last N return rate, the second part is the last N transaction volume.

This results in a large input vector (>50). I wonder if the bad results I get are related to the current selection of the input vector.

Any help on that?

Thanks! Guy

## Answer by zer0hedge (score 1)

https://quant.stackexchange.com/a/16305

50 elements input vector is actually a small one. For example, in this tutorial the size of the input vector is 784 (parameter 'nvis'). So your problem lies somewhere else.

I would recommend to start from taking these two courses on Coursera:

- Neural Networks for Machine Learning

- Machine Learning

They will provide you with some practical guidance regarding how to deal with your issue.

## Answer by Yifeng Li (score 0)

https://quant.stackexchange.com/a/22459

You can take a look at this paper: Yifeng Li, Chih-Yu Chen, Wyeth W. Wasserman: Deep Feature Selection: Theory and Application to Identify Enhancers and Promoters. RECOMB 2015: 205-217. Input features can be selected in the deep neural network by this deep feature selection model / deep lasso.

Shown in full with attribution under the source's licence. Licence: CC BY-SA 4.0 (Stack Exchange)

This summary was written by Stratmill's research agent from the original; it is not a copy of the source.