Multi-Horizon Cryptocurrency Forecasting via a Hybrid 1D-CNN-IndRNN Framework with Feature Selection Optimization

Abstract

Decentralized digital assets are characterized by pronounced volatility, non-linear dynamics, and complex temporal dependencies. These properties often render traditional econometric models and basic neural networks insufficient for long-horizon forecasting. Although volatility estimation techniques have matured, forecasting multi-horizon price movements remains a significant challenge due to non-stationarity and high-dimensional noise in these markets. To overcome these limitations, we introduce a hybrid architecture combining a One-Dimensional Convolutional Neural Network (1D-CNN) with an Independently Recurrent Neural Network (IndRNN). Designed for multi-horizon forecasting at 7, 30, and 90-day intervals, this framework processes high-dimensional cryptocurrency datasets, including market trends, technical indicators, transaction metrics, and on-chain variables. We systematically evaluate filter, wrapper, and embedded feature selection techniques (Forward, Backward, and Exhaustive) to eliminate redundancy and optimize computational cost. The 1D-CNN extracts spatial features, while the IndRNN captures long-term temporal dependencies without gradient degradation. Benchmark evaluations on Bitcoin (BTC), Ethereum (ETH), Dogecoin (DOGE), Bitcoin Cash (BCH), and IoT Chain (ITC) demonstrate the model’s effectiveness. With strict linear split validation simulating realistic future conditions, the model reduces Mean Absolute Error (MAE) by about 18\% and Mean Absolute Percentage Error (MAPE) by nearly 20\% compared to baselines such as LSTM, ANN, and SVM. In classification, the framework achieves up to 81\% accuracy on 90-day forecast horizons. Unlike resource-intensive Transformer models—which often struggle with limited data and abrupt market shifts—our lightweight architecture offers superior computational efficiency and practicality for trading, asset management, and quantitative risk modeling.