How Israel’s Ministry of Finance transformed tax revenue forecasting with AI

Tax forecasting accuracy matters. It impacts debt issuance, budget execution, fiscal planning, and the government’s ability to respond to changing economic conditions.

Israel’s Ministry of Finance wanted to strengthen the way it supports tax revenue forecasting. The Ministry worked with Amazon Web Services (AWS) Professional Services Israel to develop an automated solution that evaluates several modeling approaches, uses wider data sources, and helps economists analyze forecasting faster.

The Ministry manages approximately $160 billion in annual tax revenues across income tax, VAT, corporate tax, and other revenue streams. Even a small forecasting error can have a major financial impact. A 1% deviation may translate into roughly $1.6–$1.8 billion in additional borrowing needs and roughly $80 million in interest costs.

As explained by Dr. Lior David-Pur, senior deputy to the Accountant General and head of the Macro and Budget Division, the goal was to improve forecasting models by adding AI-based models alongside traditional statistical tools.

The Ministry’s Macro Division supports fiscal planning across key tax streams, including direct taxes, indirect taxes, and fees. Traditional models remain an important part of the professional toolkit, especially during stable periods. However, recent years have shown that economic volatility can make forecasting more difficult.

When economic conditions change quickly, models based mainly on historical patterns may take a long time to adjust. Forecasting also depends on expert knowledge that isn’t always fully documented or shared. As David-Pur explains, “Better organizational memory is something that the public sector struggles with.”

The project began with a straightforward business question: “Can we improve our forecast by using more data and better technology?”

Solution overview: A tool built for ongoing value

The Ministry and AWS Professional Services Israel built an automated solution that adds AI-based forecasting models to the existing forecasting process.

Rather than relying on one model, the solution evaluates several models and compares their performance across different tax streams and economic conditions. This matters because no single forecasting model is always the best. Some models might perform better during volatile periods. Others might be stronger when historical patterns remain stable.

“We saw the potential not just for a better model but to receive a platform that can take massive amounts of data and turn it into real business insight,” said David-Pur.

The distinction is important. A model can become outdated. A solution can evolve.

How it works

The solution is built on Amazon SageMaker with a more secure multi-account architecture that separates data storage, forecasting computation, and analytics delivery. This is an important design consideration for government environments where data governance requirements are strict.

Automated forecasting pipeline

For each tax stream (direct tax, indirect tax, and fees), the system simultaneously trains and evaluates three AI models, then automatically selects the best performer. The first model excels during volatile economic periods by adapting to new and unexpected conditions without waiting for new training data. The second performs best during periods of relative stability where established patterns hold. The third captures complex nonlinear relationships between economic indicators and tax revenues, challenging both models across volatile and stable periods alike.

The system ingests over 20 years of historical tax collection data alongside dozens of economic indicators including unemployment, gross domestic product (GDP), and government expenditure. This is a significant expansion from the limited data that the legacy approach could use.

Self-service analytics in Hebrew

The solution integrates Amazon QuickSight with generative business intelligence (BI) capabilities so Ministry officials can query forecast data using natural language, in Hebrew, without requiring technical expertise.

Instead of requesting analyst reports and waiting days for answers, Ministry officials can ask questions directly and receive visualizations immediately. For government agencies where data literacy varies widely across seniority levels, this kind of democratization has a significant impact. The people closest to policy decisions gain direct access to the information driving those decisions.

Bridging business and technology

One of the main lessons from the project was that AI in government is not only a technology project. It’s also a business project.

Forecasting requires economic judgment, fiscal understanding, knowledge of tax behavior, and familiarity with the data. Technology can support this work, but it must be connected to the real business need.

It also created value beyond the original forecasting use case. The joint work helped support reusable data pipelines and a broader data foundation that can serve additional analytical projects across the Ministry. Economic shifts are now reflected in forecasts within weeks rather than the previous 8-month lag. When conditions change, such as a new inflation trend, a shift in employment figures, or an unexpected fiscal event, forecasts update accordingly. Analysis that took several days can be done now in a few hours. This means the Ministry’s economists can focus on interpretation and policy application rather than data preparation and model maintenance.

Results: Improved accuracy and institutional knowledge capture

The Ministry reduced forecast errors by up to 40%, with projected savings of approximately $10 million monthly in optimized debt issuance. Each 1% forecast error is equivalent to approximately $80 million in annual payments.

The solution also improves organizational memory by making parts of the forecasting process more structured, reproducible, and transparent. The solution documents, tests, and reuses forecasting logic that previously depended heavily on individual experts.

The result isn’t only a faster process, it’s a more resilient one.

Lessons: Start small and define clear success criteria

Start with a clear business question. The project didn’t begin with AI for its own sake. It began with a real forecasting challenge.

Add AI to professional judgment rather than replacing it. The goal isn’t to replace economists, but to give them better tools, broader data, and faster analysis.

Build solutions, not just models

An automated platform that evaluates multiple models provides ongoing value as new algorithms emerge without requiring architectural changes. Future improvements become incremental rather than transformational, reducing long-term modernization risk.

Design for explainability from the start

Government decisions require audit trails and transparency. The Ministry’s thorough artifact storage and implementation of SHapley Additive exPlanations (SHAP) values support the accountability requirements unique to public sector applications.

What comes next

The Ministry views this solution as a foundation rather than a destination. The roadmap includes enhanced model explainability to provide transparency into which economic indicators drive predictions, integration of additional economic data sources, and expansion of the solution to support forecasting needs across other government departments. With the solution, the Ministry can produce more accurate forecasts by responding more quickly to unpredicted volatility.

For government leaders evaluating similar investments, the Ministry’s approach demonstrates how a well-designed solution can serve multiple use cases over time.

Conclusion

Israel’s Ministry of Finance is using AI to strengthen tax revenue forecasting, not to replace the professional process behind it. By combining existing models, AI based tools, broader data, and close collaboration between economists and technologists, the Ministry is building a more flexible forecasting capability.

For finance leaders evaluating similar investments, the Ministry’s experience offers both proof of concept and a practical framework. The technology is available. The methodology is proven. Better forecasting doesn’t come only from better algorithms. It comes from connecting data, technology, and professional judgment around a clear business question: “Can we improve our forecast by using more data and better technology?” David-Pur and Eibi Birman discuss the solution in more depth in this AWS video. Another example of how governments are exploring how AI powers tax transformation is this Belgian retail tax initiative, also on AWS.