BISTRO: How the BIS Is Bringing the Power of Large Language Models to Macroeconomic Forecasting

Why Macroeconomic Forecasting Is Ripe for Disruption

Traditional econometric models have served central banks and financial institutions for decades, but they come with significant limitations that are increasingly problematic in today’s volatile economic environment. Vector autoregression (VAR) models and their Bayesian counterparts require extensive specification of variable relationships, lag structures, and model forms for each forecasting task.

These bespoke approaches are not only resource-intensive but also rigid. When economic conditions change—as they have dramatically with supply chain disruptions, geopolitical tensions, and monetary policy regime shifts—traditional models often struggle to adapt. Each new forecasting challenge requires rebuilding models from scratch, creating bottlenecks for institutions that need rapid, reliable economic insights.

The recent convergence of artificial intelligence and economics presents unprecedented opportunities. Just as AI research methodologies have transformed various fields, machine learning techniques are now being successfully applied to time series forecasting. The question isn’t whether AI will transform economic analysis—it’s how quickly institutions can adapt to leverage these powerful new tools.

What Is BISTRO and Why Does It Matter?

BISTRO (BIS Time-series Regression Oracle) represents a fundamental shift in how we approach macroeconomic forecasting. Developed by researchers at the Bank for International Settlements—the “central bank of central banks”—BISTRO is what’s known as a foundational model for time series data.

The concept draws directly from the success of large language models. Just as GPT and similar models can handle diverse language tasks without being rebuilt for each application, BISTRO provides a single, general-purpose framework for macroeconomic forecasting. This isn’t just a technical improvement—it’s a paradigm shift that makes sophisticated forecasting accessible to institutions that previously lacked the resources for bespoke model development.

What makes BISTRO particularly significant is its institutional backing. The involvement of Hyun Song Shin, the BIS’s Economic Adviser and Head of Research, signals that this isn’t merely an academic exercise. It represents a strategic signal about the future of economic analysis at international financial institutions and a legitimization of AI approaches in central banking.

From Predicting Words to Predicting Economies

The technical foundation of BISTRO lies in transformer architecture—the same technology powering ChatGPT and other large language models. This might seem like an unusual connection, but the parallels are more profound than they initially appear.

“Just as LLMs do well in guessing the next word within the broader context of the sentences that precede it, foundational time series models could do well in guessing the next realisation of a macroeconomic time series within the broader context of what has been happening in the economy.”

The transformer’s attention mechanism, which allows language models to understand relationships between words across long sequences, proves equally valuable for economic data. Economic variables don’t exist in isolation—inflation affects employment, which influences consumer spending, which impacts GDP growth. BISTRO’s architecture naturally captures these complex interdependencies without requiring economists to explicitly specify every relationship.

This approach offers significant advantages over traditional methods. Rather than relying on predetermined assumptions about how economic variables interact, BISTRO learns these patterns directly from historical data. The model can identify subtle correlations and feedback loops that might be missed by conventional econometric approaches, particularly in periods of structural change.

Unconditional Forecasting: Building a Reliable Baseline

BISTRO’s unconditional forecasting capabilities provide the foundation for all other applications. These are baseline projections that don’t assume any specific future conditions—essentially asking “what is the most likely path for this economic variable given current trends?”

For central banks, reliable baseline forecasts are essential for setting monetary policy. BISTRO can generate projections for key variables like GDP growth, inflation rates, and employment levels without requiring extensive manual specification of model parameters. The system processes the BIS’s comprehensive macroeconomic database, which covers central bank statistics across more than 60 economies.

What sets BISTRO apart from traditional approaches is its ability to maintain consistency across different variables and time horizons. Traditional forecasting often requires separate models for different variables or forecast periods, leading to potential inconsistencies. BISTRO’s unified approach ensures that forecasts for related variables remain coherent and economically plausible.

The model’s performance on unconditional forecasting tasks proves competitive with established econometric benchmarks. More importantly, it achieves this performance while being significantly easier to deploy and maintain than traditional alternatives, making it particularly attractive for institutions with limited modeling resources.

Conditional Forecasting: The Real Power for Scenario Analysis

While unconditional forecasting provides valuable baseline insights, BISTRO’s conditional forecasting capabilities represent its true strategic value. This feature allows policymakers and analysts to perform sophisticated scenario analysis by asking “what if” questions about economic developments.

Consider a central bank wondering how inflation might respond if oil prices remain elevated for an extended period. Traditional approaches would require building a separate model or significantly modifying existing ones. With BISTRO, analysts can simply condition their forecast on assumed oil price paths and immediately see the implications for inflation and other variables.

“A researcher can produce a generic baseline forecast for, say, inflation and then evaluate how conditioning on different explanatory variables (and different assumptions for their evolution) modifies the baseline.”

This flexibility proves invaluable for stress testing and risk assessment. Financial institutions can explore various scenarios—geopolitical tensions, supply chain disruptions, policy changes—without rebuilding their modeling infrastructure. The speed and ease of scenario generation enables more comprehensive risk analysis and better-informed decision-making.

For regulatory bodies and international organizations like the BIS itself, conditional forecasting supports more sophisticated financial stability assessments. By rapidly testing multiple scenarios, regulators can better understand potential vulnerabilities and design appropriate policy responses.

Uncovering Hidden Nonlinearities in Economic Data

One of BISTRO’s most significant advantages lies in its ability to detect nonlinear patterns that traditional linear models miss entirely. Economic relationships are rarely linear—there are threshold effects, regime changes, and complex interactions between variables that simple linear models cannot capture.

Traditional econometric models typically assume linear relationships for computational simplicity. While economists recognize that reality is more complex, incorporating nonlinearities into traditional models requires substantial additional work and often reduces model tractability. BISTRO’s neural network architecture naturally handles nonlinear patterns without explicit specification.

This capability proves particularly valuable in periods of economic stress or structural change. During financial crises, market relationships often exhibit nonlinear behavior—small changes in conditions can trigger disproportionate responses. BISTRO can identify these patterns and incorporate them into forecasts, potentially providing earlier warning signals for emerging risks.

For example, the relationship between unemployment and inflation (the Phillips curve) has shown varying patterns over different economic periods. While linear models struggle with these regime changes, BISTRO can automatically adapt to different relationship patterns based on prevailing economic conditions. This adaptability makes forecasts more robust during periods of uncertainty.

How BISTRO Stacks Up Against Traditional Models

Benchmarking BISTRO against traditional econometric models reveals its competitive performance across key metrics. The BIS researchers compared BISTRO’s forecasting accuracy against vector autoregressions (VARs), Bayesian VARs (BVARs), and other established approaches using standard evaluation criteria.

While detailed performance metrics are available in the full working paper, the summary indicates that BISTRO performs well relative to traditional benchmarks. More importantly, it achieves this performance while offering significant operational advantages. Traditional models require extensive specification work, parameter tuning, and ongoing maintenance. BISTRO’s foundational approach dramatically reduces these requirements.

The model’s ability to handle both unconditional and conditional forecasting within the same framework represents a major practical advantage. Traditional approaches often require separate models or significant modifications for different forecasting tasks. This unified capability reduces complexity and ensures consistency across different applications.

Cost-effectiveness represents another crucial advantage. Financial technology implementations often face significant resource constraints, particularly for smaller institutions. BISTRO’s reduced requirements for specialized expertise and model maintenance make sophisticated forecasting more accessible across the financial sector.

Practical Implications for Central Banks and Financial Institutions

The practical applications of BISTRO extend across multiple areas of central banking and financial institution operations. For monetary policy committees, the model provides rapid scenario analysis capabilities that can inform policy decisions. Instead of waiting weeks for custom model runs, policymakers can explore various scenarios in real-time during meetings.

Financial stability assessments benefit significantly from BISTRO’s conditional forecasting capabilities. Regulators can quickly test how various stress scenarios might affect different sectors of the economy, enabling more comprehensive risk assessment. This rapid scenario generation supports more proactive regulatory responses to emerging threats.

Commercial banks and investment firms can leverage BISTRO for internal risk management and strategic planning. The model’s ability to generate consistent forecasts across multiple variables and time horizons supports better portfolio allocation and hedging decisions. Asset managers can explore how different economic scenarios might affect various investment strategies.

For international organizations and policy coordination efforts, BISTRO’s standardized approach facilitates better communication and comparison across different institutions. When multiple organizations use similar foundational models, their analyses become more comparable and policy discussions more productive.

Open Source by Design: Accessibility and Reproducibility

Perhaps the most democratizing aspect of BISTRO is its open-source availability. Unlike proprietary forecasting systems that require expensive licenses, BISTRO is freely available on GitHub with comprehensive documentation and pre-compiled scripts for immediate use.

The inclusion of Google Colab notebooks makes the system accessible even to institutions without significant computational infrastructure. Economists can run BISTRO forecasts directly in their web browsers without installing specialized software or managing complex technical environments. This accessibility represents a significant departure from traditional econometric tools that often require extensive technical expertise.

Open-source availability also enhances reproducibility—a crucial consideration for academic research and policy analysis. When multiple researchers can access the same tools and data, it becomes easier to verify results and build upon previous work. This transparency supports better scientific standards in economic research.

For smaller central banks and financial institutions, BISTRO’s accessibility removes significant barriers to sophisticated forecasting. Institutions that previously lacked the resources for developing in-house modeling capabilities can now access world-class forecasting tools. This democratization could significantly improve economic analysis capabilities across the global financial system.

Limitations and What BISTRO Doesn’t Do Yet

Despite its advantages, BISTRO faces several important limitations that users must understand. The model’s performance depends heavily on the quality and coverage of training data. While the BIS maintains an extensive macroeconomic database, gaps or biases in historical data will affect forecast quality.

Interpretability remains a challenge, as with many neural network approaches. Traditional econometric models often provide clear explanations for their predictions—specific coefficients indicate how changes in one variable affect another. BISTRO’s transformer architecture makes such direct interpretation more difficult, though techniques for understanding model behavior are improving.

The model’s reliance on historical patterns means it may struggle with unprecedented situations. Economic crises often involve novel combinations of factors that haven’t occurred in the training data. While BISTRO can detect complex patterns in historical data, its ability to extrapolate to truly unprecedented situations remains limited.

Computational requirements, while reduced compared to building custom models, still exceed those of simple statistical approaches. Institutions with very limited technical infrastructure may find implementation challenging, despite the Google Colab option for initial experimentation.

The Future of Economic Analysis and Foundational Models

BISTRO represents an early but significant step toward a future where foundational models transform economic analysis. Just as large language models have revolutionized natural language processing, foundational time series models could reshape how we approach economic forecasting and policy analysis.

The trend toward foundational models aligns with broader developments in artificial intelligence policy frameworks. As these models become more sophisticated and accessible, they’re likely to influence not just technical analysis but also policy-making processes and institutional structures.

Future developments might include specialized foundational models for different economic domains—financial markets, labor economics, international trade—each building on BISTRO’s foundational approach. Cross-domain integration could eventually enable comprehensive economic system modeling that surpasses traditional approaches in both accuracy and scope.

The open-source nature of BISTRO suggests a collaborative future for economic modeling. Rather than institutions developing proprietary systems in isolation, the field might move toward shared foundational infrastructure with specialized applications. This collaborative approach could accelerate innovation and improve overall modeling quality across the financial sector.