Forecasting cryptocurrencies prices using data driven level set fuzzy models

Abstract

The paper develops fuzzy models to forecast cryptocurrencies prices using a data-driven fuzzy modeling procedure based on level set. Data-driven level set is a novel fuzzy modeling method that differs from linguistic and functional fuzzy modeling in how the fuzzy rules are built and processed. The level set-based model outputs the weighted average of output functions of active fuzzy rules. Output functions map the activation levels of the fuzzy rules directly in model outputs. Computational experiments are done to evaluate the level set method in one-step-ahead forecasting of the closing prices of cryptocurrencies. Comparisons are made with the autoregressive integrated moving average, multi layer neural network, and the naïve random walk as a benchmark for Cardano, Binance Coin, Bitcoin, Ethereum, Chainlink, Litecoin, Tron, Stellar, Monero and Ripple. The results suggest that the random walk outperforms most methods addressed in this paper, confirming the Meese–Rogoff puzzle for the case of digital coins, i.e. the difficulty to surpass the naïve random walk in predicting exchange rates. However, when performance is measured by the direction of price change, the level set-based fuzzy modeling performs best amongst the remaining methods.