The Central Bank of The Bahamas Automates its Forecasting Model with the Anaconda Platform

Every forecast cycle, the Research Department (the Department) at the Central Bank of The Bahamas (the Central Bank) has to answer the same question: where does the Bahamian economy stand, and where is it headed? For an economy pegged to the U.S. dollar, that answer depends on tracking dozens of indicators, from GDP and inflation to reserves and financial-sector health, and modelling how external shocks ripple through it. For years, that work ran through an Excel-based Financial Programming Model (FPM), hand-fed with data pasted from international sources each cycle. Today, the Department runs a fully coded forecasting system built on Python and the Anaconda Platform that automates the same trusted economics, cuts a multi-day forecast round down to an afternoon, and gives every economist in the department a shared, auditable view of the economy.

The team needed to leverage cutting-edge machine learning techniques, but their security and compliance requirements demanded strict governance over code dependencies in production systems. This created a fundamental tension: how to move fast with innovation while maintaining the rigorous security standards banking regulators require.

“With AI getting more complex, having loads of disparate data inputs and making sure that the algorithms we deploy are stable and perform as we expect is quite difficult,” explains the Head of Data Science, who oversees a team responsible for developing algorithms across fraud detection, credit risk, and regulatory compliance. “The only way to do that is to use data science. You have to use machine learning to deliver anything that is sufficiently complex.”

The Challenge: A Trusted Model, Built in a Workbook​

The Central Bank’s underlying FPM, the International Monetary Fund (IMF) style framework central banks use to translate economic assumptions into policy scenarios, had served the Department well for years. The workbook it ran took time to produce results; thereby, delaying the associated work streams connected to them. 

Each forecast round meant manually pasting data from sources like the IMF’s International Financial Statistics (IFS), World Economic Outlook (WEO) releases and the World Bank’s data, into a spreadsheet held together by cross-sheet links that could break silently. As Dr. Allan Wright, Manager in the Department at the Central Bank, described the prior process: “The Financial Programming Model is a useful tool, which has served our department well. However, the process was particularly time-consuming. When you’re using spreadsheets, especially ones with a lot of macros and equations, they can get hung up on you very fast, and you can lose the connections between different sheets in the workbook.”

That Excel based framework created compounding risk:

• Manual, repetitive data entry from multiple external sources every forecast cycle, with an audit trail showing where a number came from.
• Imperfect cross-sheet logic, where a single broken cell could propagate errors silently through the model.
• Slow, time consuming-to-reproduce scenario runs, with a full round taking days. 
• Concentrated institutional knowledge, with the model’s mechanics understood mainly by whoever maintained the file, making the process hard to scale or hand off.

The Approach: Coding the Model, Not Just the Spreadsheet​

Rather than replace the FPM’s underlying economics, the Department’s team rebuilt its mechanics in Python, running on the Anaconda Platform, to keep the model’s logic version-controlled, tested, and reproducible across the Department rather than dependent on a single workbook.

The resulting platform is built in two connected layers. A data pipeline pulls from four official Application Programming Interfaces—the World Bank, IMF WEO, IMF SDMX (Statistical Data and Metadata eXchange), and FRED (Federal Reserve Economic Data)—to automatically refresh a registry of 257 economic indicators across 14 categories, with every value tagged to show which source supplied it. A set of model engines then runs the same nine-step financial-programming logic the Central Bank has always used, from measuring output gaps through fiscal, external, and banking-sector blocks, to a tested codebase instead of a chain of spreadsheet formulas. A second engine, a Python implementation of the European Central Bank’s Nowcasting Toolbox, gives the Department a live estimate of current-quarter GDP ahead of official data releases.

Building on the Anaconda Platform gave the team a consistent, governed Python environment to develop and maintain that codebase; a trusted package ecosystem to draw on for the statistical and data-pipeline work; and a foundation the entire Department, not just one analyst, can run, test, and extend.

“We have real time, at any point in time, information on all of the key variables that we have,” Dr. Wright said. “At the same point in time, we are now able to do advanced analysis, as well as forecasting on all of these variables.”

The Results

The new system has changed the speed of the Department’s forecasting work:

• From days to an afternoon. The mechanical work of a full forecast round, once measured in days, now runs in an afternoon, with a full data refresh completing in about 20 seconds and a multi-scenario projection running in seconds.
• 257 indicators, fully sourced. Every one of the model’s 257 indicators, across 14 categories, is pulled from official sources with a recorded audit trail showing exactly where each value came from.
• Zero accounting failures. Fifteen national-accounting identities are re-checked automatically on every model run, currently with zero failures, replacing manual, late-night verification.
• Near real-time visibility. A dedicated nowcasting model combines three independent modelling approaches to give the Department a current, up-to-date read on economic conditions ahead of official data releases, rather than waiting on a quarterly reporting cycle.
• A shared, testable model. Documentation mapping every mechanism back to the original workbook means the model is no longer tied to any one analyst’s knowledge.

“We have real time, at any point in time, information on all of the key variables that we have. At the same point in time, we are now able to do advanced analysis, as well as forecasting on all of these variables.” — Dr. Allan Wright, Manager, Research Department, Central Bank of The Bahamas

Looking Forward

The Department’s roadmap extends the platform in several directions: GDP-by-industry detail, richer monitoring of data readiness, and additional model families, including benchmark time-series and cross-country comparisons.

The Department’s ambitions also go beyond the current system’s scope. The team wants to build similar rigor into tracking residential and commercial asset prices, including modelling how hurricane exposure and climate risk could affect property values and insurance costs across the islands, work tied to the multilateral climate financing the Bahamian government receives. A related effort would track the country’s evolving tourism and cruise-passenger economics, as the visitor mix shifts toward a younger, higher-frequency segment with different spending patterns.

Longer term, Dr. Wright described a vision of the platform where researchers, students, other central banks, and the public could view live Bahamian economic data and run their own scenarios directly, without contacting the Central Bank.