Expert System for Bitcoin Forecasting: Integrating Global Liquidity via TimeXer Transformers

Expert System for Bitcoin Forecasting: Integrating Global Liquidity via TimeXer Transformers Sravan Karthick T∗ RV College of Engineering (RVCE), Bengaluru, India Rakshita A RV College of Engineering (RVCE), Bengaluru, India Dr. Minal Moharir RV College of Engineering (RVCE), Bengaluru, India Abstract Bitcoin price forecasting is characterized by extreme volatility and non-stationarity, making it hard for univariate time-series prediction models over long-horizons. By integrating Global M2 Liquidity as an exogenous variable, aggregated from 18 major economies, we improve the predictions over long-horizons. Using TimeXer, we compare a liquidity-conditioned univariate forecasting model (TimeXer- Exog) against state-of-the-art univariate models, Long Short-Term Memory(LSTM), Neural Basis Expansion Analysis for interpretable Time-Series (N-BEATS), Patch Time-Series Trans- former(PatchTST), and standard TimeXer. Experiments conducted on daily Bitcoin price data from January 2020 to August 2025 demonstrate that explicit macroeconomic conditioning sig- nificantly stabilizes long-horizon forecasts. At a 70-day forecast horizon, the proposed TimeXer-Exog model achieves a Mean Squared Error (MSE) of 10.814, scaled by 107, which outperformed the univariate TimeXer baseline by over 89%. The results show that explicitly conditioning deep learning models on Global Liquidity as an exogenous variable gives substantial improvements in long-horizon Bitcoin price forecasting. Keywords: Bitcoin forecasting, Global liquidity, TimeXer, Transformers, Deep learning, Macroe- conomic conditioning, Long-horizon forecasting 1 Introduction Early attempts to forecast Bitcoin prices were made using statistical tools, especially the ARIMA models (Sagheer et al., 2025; Xu, 2025). ARIMA is good at picking up short-term linear patterns, like the ones in a relatively calm equity index over a few weeks. However, bitcoin does not behave like that. Its price is noisy and very volatile, which statistical models and tools do not capture. ARIMA requires stationarity and linear structure, which is exactly what bitcoin lacks. Large rallies, sudden crashes, or regime shifts often get smoothed away (Pratas et al., 2023; Mizdrakovic, 2024). ∗Corresponding author. Email: sravankt.cs20@rvce.edu.in 1 arXiv:2512.22326v2 [cs.LG] 11 Jan 2026 As a result, these models can perform well over a short horizon, their accuracy drops once the forecast horizon is about 60 or 90 days (Mousa, 2025; Kareem and Aue, 2025). Researchers moved to RNN models, which were designed to handle sequences. Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU) models were successful in learning the non- linear patterns from the data (Bagheri and Giudici, 2025; Xu, 2025). They can, for example, react differently to a slow upward trend than to a sudden spike triggered by a regulatory headline. These models are not perfect. Because they process data step by step, they scale poorly and struggle when dependencies stretch far back in time. Capturing relationships across hundreds of days still remains hard, and unreliable (Zhao et al., 2022; Zheng, 2025). More recently, attention has shifted toward Transformer-based models, time series forecasting moved to self-attention rather than strict sequence processing (Nie, 2022; Zhao et al., 2022). Instead of remembering the past one step at a time, these models can look across an entire window at once. PatchTST takes this idea further by breaking the time series into patches, small segments that preserve local structure while keeping computation manageable (Nie, 2022). For time series fore- casting which require long look-back windows, Transformer models have repeatedly outperformed LSTM-based approaches, especially for long-horizon predictions (Nie, 2022; Zheng, 2025). Long-term forecasting improves when models are allowed to look beyond price history itself. The TimeXer framework reflects this shift. It combines patch-wise attention for the main series with cross-attention for external variables, making it possible to integrate data that arrive at different frequencies and are not neatly aligned in time (Wang, 2024). This is closer to how real financial systems operate. Macroeconomic indicators do not update on a daily basis, yet markets tend to respond to them gradually and unevenly. Global liquidity is often measured by M2 money supply, which has particularly attracted atten- tion (Gu and Chen, 2025). Although the relationship is not perfectly stable, the logic is intuitive. According to empirical work using time-varying Granger causality, the influence of M2 on Bitcoin prices is not constant. It strengthens during expansionary regimes and weakens in other regimes (Gu and Chen, 2025). Co-integration analyses go further by pointing towards a long-run equi- librium relationship in which Bitcoin prices respond more than proportionally to liquidity growth (Kokabian, 2025). It is important to note that these effects do not come into the pi

This content is AI-processed based on open access ArXiv data.