Predicting Bitcoin’s Future: Global Money Supply as a Key Indicator

Author: Denis Avetisyan

A new expert system leverages global liquidity data alongside advanced time-series forecasting to achieve more accurate long-term Bitcoin price predictions.

The TimeXer architecture integrates both internal and external data through dedicated embedding and attention mechanisms, enabling a nuanced understanding of temporal dynamics.

Integrating M2 money supply into a TimeXer Transformer model significantly improves long-horizon forecasting performance compared to price-only approaches.

Predicting Bitcoin’s volatile price movements remains a persistent challenge for conventional time-series models. This is addressed in ‘Expert System for Bitcoin Forecasting: Integrating Global Liquidity via TimeXer Transformers’, which proposes a novel forecasting system incorporating global M2 liquidity as an exogenous variable within a TimeXer Transformer architecture. Results demonstrate that explicitly conditioning the model on macroeconomic factors substantially improves long-horizon prediction accuracy, achieving an 89% reduction in mean squared error compared to a univariate baseline. Could this approach unlock more stable and reliable forecasting for other volatile assets influenced by global economic conditions?

The Limits of Prediction: Why Traditional Methods Fall Short

Initial forays into Bitcoin price prediction frequently relied on models like the Autoregressive Integrated Moving Average (ARIMA), a statistical approach traditionally successful with stable time series data. However, these early attempts quickly revealed fundamental limitations when applied to the cryptocurrency market. Bitcoin’s price action is characterized by extreme volatility and non-linear dynamics – meaning past patterns are a poor predictor of future behavior, and relationships between data points aren’t consistent. The ARIMA model, designed for stationary data, struggled to adapt to these conditions, producing forecasts with limited accuracy and failing to capture the rapid, often unpredictable, price swings inherent in the emerging digital asset. This highlighted the need for more sophisticated methodologies capable of handling the unique complexities of the Bitcoin market.

Recurrent Neural Networks (RNNs), initially promising for Bitcoin price prediction due to their ability to learn temporal patterns, encountered significant obstacles when forecasting over extended periods. The core limitation stems from the ‘vanishing gradient’ problem, where signals representing early data points weaken as they propagate through the network during training, hindering the model’s capacity to establish connections between distant events in the time series. Consequently, RNNs struggle to capture long-range dependencies – crucial for understanding how past price action influences future values – leading to diminished accuracy in long-term forecasts. While adept at recognizing immediate trends, the network’s ‘memory’ effectively fades when predicting further into the future, highlighting the need for more sophisticated architectures capable of preserving information across longer sequences.

Bitcoin's open price correlates with global liquidity, exhibiting a leading indicator relationship of approximately 12 weeks after changes in liquidity. — Bitcoin’s open price correlates with global liquidity, exhibiting a leading indicator relationship of approximately 12 weeks after changes in liquidity.

Advancing the State of the Art: The Rise of Recurrent Networks

Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU) models were developed as improvements to traditional Recurrent Neural Networks (RNNs) to mitigate the vanishing gradient problem encountered when processing long sequences. Standard RNNs struggle to learn dependencies between data points separated by many time steps because gradients used during training diminish exponentially as they are backpropagated through time. LSTM and GRU architectures introduce gating mechanisms – specifically, input, forget, and output gates in LSTMs, and update and reset gates in GRUs – that regulate the flow of information, allowing the network to retain relevant data over extended periods. These gates utilize sigmoid and tanh functions to control which information is passed through, stored in the cell state (LSTM) or hidden state (GRU), and ultimately outputted, thereby enabling the models to more effectively capture long-term dependencies in sequential data such as time series or natural language.

While Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU) models represent advancements over traditional Recurrent Neural Networks in capturing long-term dependencies, limitations persist when applied to Bitcoin price prediction. These models still encounter difficulties with extremely long sequences of data, potentially losing crucial information from earlier time steps due to computational constraints and memory limitations. Furthermore, effectively integrating external factors – such as macroeconomic indicators, regulatory news, or social media sentiment – remains a challenge. Simply appending these factors as additional input features often proves insufficient, as the models struggle to discern the complex and non-linear relationships between these external variables and Bitcoin price movements. This restricts their ability to provide accurate predictions, particularly during periods of high volatility or significant external influence.

A Paradigm Shift: The Transformer Architecture for Bitcoin Forecasting

The Transformer architecture, initially developed for natural language processing, addresses limitations of recurrent neural networks (RNNs) in time series forecasting through its core self-attention mechanism. Unlike RNNs which process sequential data linearly, Transformers analyze the entire input sequence simultaneously, enabling the model to directly assess the relationships between any two data points, regardless of their distance. This capability is crucial for identifying long-range dependencies – patterns spanning extended periods – that significantly influence future values. Furthermore, the self-attention mechanism allows for parallel computation, drastically reducing processing time compared to the sequential nature of RNNs. This parallelization, combined with the ability to capture complex temporal relationships, results in demonstrably improved performance in long-term forecasting tasks, particularly when dealing with volatile and complex datasets like cryptocurrency price movements.

PatchTST improves Transformer-based time series forecasting by segmenting the input data into smaller, non-overlapping patches. This approach reduces the computational complexity associated with the self-attention mechanism, particularly for long sequences, by limiting the number of pairwise comparisons. Instead of attending to every element in the time series, attention is focused within and between these patches, significantly decreasing the quadratic complexity of standard Transformers from O(n²) to a more manageable level. This allows for the processing of longer time horizons and facilitates improved accuracy by enabling the model to better capture local and global patterns within the data.

TimeXer is a forecasting framework designed to incorporate external variables, specifically global M2 liquidity, into Bitcoin price predictions. It utilizes a Cross-Attention mechanism, allowing the model to weigh the relevance of exogenous data when generating forecasts. Evaluations demonstrated that the integration of global M2 liquidity via TimeXer resulted in an 89% reduction in forecast error for 70-day Bitcoin price predictions when compared to established benchmark models. This indicates a substantial improvement in predictive accuracy through the inclusion of macroeconomic indicators within the forecasting process.

Unveiling the Underlying Equilibrium: Global Liquidity and Bitcoin Price

Cointegration analysis demonstrates a statistically significant long-run equilibrium relationship between Bitcoin prices and global liquidity, indicating these two time series move together over time, despite potential short-term fluctuations. This implies that global liquidity, representing the total amount of easily accessible money and credit in an economy, is a fundamental driver of Bitcoin’s price. The existence of cointegration suggests that deviations from this equilibrium are temporary and will eventually revert, allowing for potential predictive modeling based on macroeconomic indicators. This relationship does not imply causation, but establishes a predictable connection suitable for quantitative analysis within the TimeXer framework.

The Error Correction Term (ECT) in the cointegration analysis quantifies the rate at which Bitcoin prices and global liquidity revert to their long-run equilibrium following a shock. An estimated ECT of -0.12 indicates that approximately 12% of any temporary deviation from this equilibrium is corrected each month. This implies a relatively rapid, though not instantaneous, adjustment process; deviations do not persist indefinitely but are systematically reduced over time, reinforcing the observed long-run relationship between the two variables. The negative sign confirms that the correction acts to restore the equilibrium; a positive deviation will be offset by a decrease in Bitcoin price, and vice-versa.

The TimeXer model incorporates M2 money supply as a key exogenous variable to improve Bitcoin price prediction. Long-Run Elasticity analysis demonstrates a 2.65 relationship between M2 money supply and Bitcoin price, meaning a 1% increase in the M2 money supply is associated with a 2.65% increase in Bitcoin price. This suggests a substantial sensitivity of Bitcoin’s price to changes in global liquidity, as measured by M2, and highlights the variable’s importance in understanding the cryptocurrency’s market dynamics within the model.

Validating Forecasts and Establishing Confidence in Predictive Models

The Model Confidence Set offers a robust statistical method for gauging the reliability of forecasting models, such as TimeXer, by rigorously assessing which models consistently generate accurate predictions. This technique doesn’t simply rank models; instead, it identifies a subset that performs comparably well, effectively establishing a benchmark of consistently reliable options. By employing statistical testing, the Model Confidence Set determines if a model’s performance is statistically indistinguishable from the best-performing models within the evaluated group, providing a confidence interval around the forecast. This approach is particularly valuable in dynamic fields like Bitcoin price prediction, where identifying consistently accurate models is crucial for informed decision-making and effective risk management. Ultimately, the Model Confidence Set moves beyond simple point forecasts, offering a quantifiable measure of forecast reliability and bolstering confidence in predictive analytics.

Quantifying the reliability of Bitcoin price forecasts is now achievable through rigorous model confidence assessment, offering a substantial benefit to investors and risk managers. Traditionally, forecasting accuracy has been evaluated through point predictions, but this approach fails to capture the uncertainty inherent in financial markets. By employing techniques like the Model Confidence Set, the probability of a model consistently delivering accurate predictions can be determined. This allows for a more nuanced understanding of forecast trustworthiness, moving beyond simple accuracy scores to provide actionable insights. Investors can then utilize these confidence levels to adjust portfolio allocations, refine trading strategies, and establish more informed risk parameters, ultimately leading to better-capitalized decisions in the volatile cryptocurrency landscape.

Rigorous validation procedures confirm the capacity of Transformer-based models to substantially enhance long-term forecasting accuracy, particularly when integrated with relevant exogenous variables. The TimeXer-Exog model, subjected to Model Confidence Set analysis, achieved a P-value of 1.000 at both 63- and 70-day forecasting horizons. This statistically significant result-indicating the model consistently outperformed all others within the tested ensemble-demonstrates a clear advantage in predicting future values. The findings suggest that incorporating external factors into the Transformer architecture allows for more robust and reliable long-term forecasts, potentially offering valuable insights for decision-making in dynamic systems.

The pursuit of predictive accuracy, as demonstrated by the integration of global liquidity indicators into the TimeXer Transformer model, echoes a fundamental principle of efficient design. This research exemplifies how distilling complex, external factors-like M2 money supply-into a concise and relevant input significantly refines forecasting capabilities. Ada Lovelace observed, “The Analytical Engine has no pretensions whatever to originate anything. It can do whatever we know how to order it to perform.” The model, much like the Engine, operates on provided directives; however, the power lies not in inherent creation, but in the clarity and relevance of the information supplied. This paper shows that a lossless compression of global economic signals translates directly into a more insightful forecast, minimizing extraneous noise and maximizing predictive power.

Where Do We Go From Here?

The demonstrated improvement in forecasting accuracy, achieved through the integration of global liquidity indicators, does not represent an arrival. It exposes the prior state as unnecessarily complex. Models previously burdened with elaborate architectures, attempting to divine future prices from past prices alone, now appear…indulgent. The signal was not within the price history, but external to it, a simple truth obscured by a preference for internal logic.

The immediate task is not further refinement of the TimeXer architecture, but a ruthless pruning. What layers, what attention heads, truly contribute after the introduction of M2 supply as a primary input? The expectation is a significant reduction in complexity, a shedding of algorithmic weight. Furthermore, the singular focus on M2 represents a convenient, not necessarily optimal, solution. The broader question remains: what other, more parsimonious, macroeconomic indicators hold predictive power, and how can these be integrated without reintroducing the bloat?

Ultimately, the pursuit of perfect prediction is a fool’s errand. The aim should be sufficient prediction – a model that reliably captures the dominant forces, and then stops. The elegance lies not in anticipating every fluctuation, but in acknowledging the inherent noise, and building a system that respects its limitations. The future of this field is not more data, or more layers; it is a commitment to subtraction.

Original article: https://arxiv.org/pdf/2512.22326.pdf

Contact the author: https://www.linkedin.com/in/avetisyan/