Evaluating and Finetuning Models For Financial Time Series Forecasting

Tom Djaaleb, Julien Romero

Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation learning

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Keywords: time series forecasting, finance, metrics

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.

TL;DR: We study the topic of financial time series forecasting, show it is complex, and was pooly treated by previous works.

Abstract: Time series forecasting is a challenging task as it is subject to a lot of noise, and the predictions often depend on external events. Still, recent deep learning techniques advanced the state-of-the-art on certain datasets, while they keep failing on other noisy datasets. This paper studies the case of financial time series forecasting, a problem that exhibits both a high noise and many unknown dependencies. We will show that the current evaluation pipelines are imperfect and forget a trivial baseline that can beat most models. We propose a new evaluation pipeline that is better suited for our task, and we run this pipeline on recent models. This pipeline is based on the idea of deciding which assets to buy and sell rather than predicting exact prices. Next, as the small datasets used in current approaches limit the size of the models, we train a general model on a massive dataset (containing a hundred times more data points than existing datasets) and show this model can be finetuned to improve the performance on small datasets. All our code and models will be published to help the community bootstrap and evaluate their future models.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 6351