DUAL-MODEL BIDIRECTIONAL LSTM FOR CRYPTOCURRENCY PRICEPREDICTION: BETTER ACCURACY THROUGH COMPLEMENTARY TIME SCALES

Abstract

Predicting cryptocurrency prices is hard because markets show both long-term trends and sudden short-term changes. Most prediction systems use a single model that either looks at years of history (missing recent changes) or focuses only on recent data (missing established patterns). We developed a new approach using two LSTM networks working together. One model train on a longer year of data to learn lasting patterns. The other trains on just the last two months to catch current trends. A smart system combines their predictions based on how volatile the market is right now. We tested this on Bitcoin, Ethereum, and Cardano from 2022 to 2025. The results were impressive. Bitcoin prediction errors dropped from 3.4% to 1.5% - that's 56% better. Ethereum improved 34% and Cardano improved 37%. On average across all three, we got 42% better accuracy than the previous best method. The system trains in just 27 seconds, so it can update daily to stay current. Trading based on these predictions would be profitable 85- 87% of the time. The confidence scores the system gives (88-92%) match actual accuracy within 1%, meaning traders can trust them for managing risk. This dual-model approach works because it keeps both perspectives active - you don't have to choose between learning from history and adapting to current conditions.