Machine Learning Models for Financial Time Series Classification
Summary
The document discusses model choices for classifying financial time series into positive, negative, or neutral outcomes according to whether a forward return crosses a chosen threshold. The responses point to support vector machines and random forests as commonly used approaches, alongside neural network architectures. They also mention recurrent neural networks, especially LSTMs, and reinforcement learning for portfolio-oriented applications.
One answer reports that LSTMs appeared in just over 40% of a survey of research papers since 2018, and that many surveyed papers tested only one algorithm. Another says the author had found more success with random forests and SVMs than with sequential models, while describing LSTMs as potentially effective for suitable problems and datasets but resource-intensive. These observations describe literature prevalence and individual experience; they do not establish that one model will outperform others for a particular asset, feature set, or labeling scheme. Model choice still depends on the task and data, and the document offers references rather than a controlled comparison of methods.
Key ideas
- The example frames financial forecasting as three-class classification based on thresholded forward returns.
- Support vector machines, random forests, and neural networks are among the model families discussed.
- A survey cited in the document found LSTMs common in recent financial time-series papers.
- LSTMs may require substantial resources, and their suitability depends on the problem and dataset.
- Reported model popularity or individual success does not establish which method will work best on a new dataset.
Tags
Full text
# What machine learning method is more suitable for prediction of financial time series? # What machine learning method is more suitable for prediction of financial time series? I have time series data for various assets and which I transform to create various features. I have framed the problem as a classification task where I attempt to predict either a positive or negative move above some threshold. Example: If the next n day returns are above the threshold then label as a one. If below the threshold then a -1. If the threshold is not reached then a 0. (This is inline with much of what I see in the literature) At this stage I have only investigated an SVM classifier but I am wondering if there are more appropriate models to use? If someone familiar with the literature could please point me to a few models that are common in forecasting financial time series, that would be great! One idea I have had is the HM-SVM model architecture, am I getting warmer? Are there specific model tricks? ## Answer by Jacques Joubert (score 5) https://quant.stackexchange.com/a/44930 From what I have read, there are 3 popular algorithms for financial time series. Random Forests and SVMs, then followed by Neural Network Architectures. There are a couple of good papers, to name a few: - Empirical Asset Pricing via Machine Learning - Deep neural networks, gradient-boosted trees, random forests: Statistical arbitrage on the S&P 500 - An Empirical Comparison of Machine Learning Models for Time Series Forecasting - Evaluating multiple classifiers for stock price direction prediction - Predicting stock returns by classifier ensembles - Predicting stock market index using fusion of machine learning techniques I have a lot of hope for the sequential models such as RNNs but I have had more success with Random Forests and SVMs. ## Answer by Oeyvind (score 4) https://quant.stackexchange.com/a/68412 I wrote a masters thesis related to machine learning in finance, and during this process I surveyed about 200 of the research papers that were written about the topic since 2018. This is the distribution of the algorithms used in the research papers: LSTM is by far the most used machine learning algorithm used to predict financial time series and used in just above 40% of the surveyed papers. It seems like in both research articles as well as in code examples provided on different websites, LSTM is the go-to algorithm for predicting financial time series. LSTM usually gives excellent results when paired with the correct problems and datasets, although it is resource intensive. Another contender is Reinforcement learning, which is more popular when it comes to financial portfolios. 58% of the papers focused on only one algorithm, while 18% used two. 12% used 3 and 2% used 4. None used more than 4 different models.
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.