Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Leveraging Time Series Categorization and Temporal Fusion Transformers to Improve Cryptocurrency Price Forecasting

Arash Peik, Mohammad Ali Zare Chahooki, Amin Milani Fard, Mehdi Agha Sarram

TL;DR

This work tackles cryptocurrency price forecasting under data-scarce conditions by proposing a categorization of financial time series into subseries that exhibit similar behavior. Each subseries category is modeled with a category-specific Temporal Fusion Transformer (TFT), and a probabilistic selector (e.g., Markov chain or LSTM) routes predictions to the appropriate TFT. By augmenting training data through cross-cryptocurrency pooling and leveraging the TFT's attention-based forecasting with time-varying covariates, the approach achieves higher accuracy and profitability than baselines such as LSTM or single TFT models. The study demonstrates that even modest improvements in the selector’s accuracy can yield substantial gains in trading profitability, with potential applicability to other financial domains. Overall, the method offers an adaptive, data-efficient framework for financial market prediction that can adjust to changing subseries dynamics in real time.

Abstract

Organizing and managing cryptocurrency portfolios and decision-making on transactions is crucial in this market. Optimal selection of assets is one of the main challenges that requires accurate prediction of the price of cryptocurrencies. In this work, we categorize the financial time series into several similar subseries to increase prediction accuracy by learning each subseries category with similar behavior. For each category of the subseries, we create a deep learning model based on the attention mechanism to predict the next step of each subseries. Due to the limited amount of cryptocurrency data for training models, if the number of categories increases, the amount of training data for each model will decrease, and some complex models will not be trained well due to the large number of parameters. To overcome this challenge, we propose to combine the time series data of other cryptocurrencies to increase the amount of data for each category, hence increasing the accuracy of the models corresponding to each category.

Leveraging Time Series Categorization and Temporal Fusion Transformers to Improve Cryptocurrency Price Forecasting

TL;DR

This work tackles cryptocurrency price forecasting under data-scarce conditions by proposing a categorization of financial time series into subseries that exhibit similar behavior. Each subseries category is modeled with a category-specific Temporal Fusion Transformer (TFT), and a probabilistic selector (e.g., Markov chain or LSTM) routes predictions to the appropriate TFT. By augmenting training data through cross-cryptocurrency pooling and leveraging the TFT's attention-based forecasting with time-varying covariates, the approach achieves higher accuracy and profitability than baselines such as LSTM or single TFT models. The study demonstrates that even modest improvements in the selector’s accuracy can yield substantial gains in trading profitability, with potential applicability to other financial domains. Overall, the method offers an adaptive, data-efficient framework for financial market prediction that can adjust to changing subseries dynamics in real time.

Abstract

Organizing and managing cryptocurrency portfolios and decision-making on transactions is crucial in this market. Optimal selection of assets is one of the main challenges that requires accurate prediction of the price of cryptocurrencies. In this work, we categorize the financial time series into several similar subseries to increase prediction accuracy by learning each subseries category with similar behavior. For each category of the subseries, we create a deep learning model based on the attention mechanism to predict the next step of each subseries. Due to the limited amount of cryptocurrency data for training models, if the number of categories increases, the amount of training data for each model will decrease, and some complex models will not be trained well due to the large number of parameters. To overcome this challenge, we propose to combine the time series data of other cryptocurrencies to increase the amount of data for each category, hence increasing the accuracy of the models corresponding to each category.

Paper Structure

This paper contains 13 sections, 5 figures, 2 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Work
  3. The Proposed Approach
  4. Data Preprocessing
  5. Time Series Categorization
  6. Probabilistic Model Selector
  7. Temporal Fusion Transformer Construction
  8. Experiments and Results
  9. Dataset
  10. Implementation and Settings
  11. Comparison
  12. Discussion
  13. Conclusion and Future Work

Figures (5)

  • Figure 1: Overview of our proposed method.
  • Figure 2: Cluster with different behavior.
  • Figure 3: Time subseries of 7 consecutive moments of the Bitcoin-Tether currency pair
  • Figure 4: Two subseries with similar appearance, same acceleration, but different behavior.
  • Figure 5: Candlestick chart of two subseries with the same acceleration.