New AI Skill Turns Agents Into Macroeconomists

Overview

A new open-source installation for AI agents introduces macroeconomic nowcasting capabilities, allowing large language models to function as structural economists rather than simple data retrievers. The project, titled ai-economist-skill, is designed to bridge the gap between advanced machine learning techniques and established economic theory, specifically targeting GDP forecasting and central bank policy diagnostics.

Rather than treating AI as a black-box predictor that hallucinates correlations, this framework positions machine learning as an auxiliary calibration layer. It enhances traditional structural models by identifying where those models may be missing critical information, thereby providing agents with economist-style interpretation and rigorous validation metrics.

The skill does not merely point the user to a Python command; it drives a structured output contract that includes forecast target periods, data-through dates, baseline estimates, and driven-factor decompositions.

Architectural Philosophy

The core innovation of this release lies in its refusal to treat ML as the primary forecaster. In traditional predictive analytics, deep learning models often obscure the "why" behind a prediction. This skill prioritizes transparency and causal inference, ensuring that the AI agent’s output is interpretable and grounded in economic logic.

The system operates on a dual-layer architecture. The first layer relies on bridge-equation baselines, which are standard econometric models used across central banks and research institutions. These baselines provide the initial forecast based on historical relationships between macro indicators. The second layer employs ML for bounded residual calibration. If the structural model deviates from recent observable trends, the ML component adjusts the forecast, but only within calibrated bounds that prevent erratic or unrealistic projections.

This approach ensures that the agent remains faithful to economic theory while accounting for real-time market anomalies. The output is not a hidden predictor but a structured report detailing validation notes, limitations, and the specific factors driving the current economic outlook.

Key Modules and Capabilities

The skill encompasses three primary functional areas, each designed to serve a distinct role in macroeconomic analysis.

GDP Nowcasting

The nowcasting engine utilizes a structural bridge model with SVD factor extraction to synthesize disparate economic data points into a coherent quarterly GDP estimate. This module is particularly notable for its handling of "ragged-edge" data—situations where recent monthly observations are missing or incomplete.

For the United States, the model incorporates high-frequency auxiliary variables, including:

• Initial jobless claims
• Housing market indicators
• Durable goods orders
• Real disposable income
• Financial conditions indices
• Yield curve spreads

For Canada, the system leverages Statistics Canada public CSV tables for retail sales, alongside CPI, CAD/USD exchange rates, WTI oil prices, and US demand spillover measures. The agent presents both the baseline estimate and the ML-calibrated result side by side, allowing users to assess the magnitude of the adjustment.

Policy Rate Diagnostics

Moving beyond growth forecasts, the skill provides diagnostics for central bank policy. It calculates readings based on the Taylor Rule, offering both the classic 1993 and 1999 variants, as well as nonlinear inflation-response models.

The analysis includes:

• Base Taylor results: The main structural signal derived from standard output-gap proxies.
• Data-Enhanced Taylor results: A secondary signal that learns historical residual adjustments from activity, inflation pressure, financial conditions, external pressure, and labor market cooling.

The enhancement decomposition is reported in percentage-point contributions, giving users insight into which macro factors are currently pushing the recommended policy rate up or down.

Validation and Backtesting

To prevent overfitting and ensure reliability, the skill includes an expanding-window backtest engine. This tool compares baseline versus calibrated out-of-sample performance, filtered by release lags to avoid look-ahead bias. The agent reports R-squared and RMSE values for both models on the same validation window.

Crucially, the system discloses revised-data limitations. It does not treat vintage data as ground truth but acknowledges when historical revisions might affect the accuracy of retrospective analysis.

Installation and Usage

The project is distributed as an installable agent skill, primarily housed within the economics-ml/ directory. The repository-level dashboard/, tests/, and assets/ folders are supporting files for development and are not required for standard agent use.

Users can install the skill on macOS or Linux by cloning the repository and copying the economics-ml/ folder into their local skills directory, followed by installing Python dependencies via pip. Windows PowerShell users can achieve the same result using specific path manipulation commands to copy the skill bundle.

After installation, users can prompt the agent with natural language queries such as: "Use the Economics ML skill to explain the latest US and Canada GDP nowcast drivers" or "Audit whether the backtest has forward-information leakage."

For those interested only in structured interpretation without live data pulls, the SKILL.md file alone suffices. However, accessing real-time FRED API data and running backtests requires the full runtime bundle and the configuration of a local FRED_API_KEY.

Data sources include the Federal Reserve Economic Data (FRED) API, Bureau of Labor Statistics (BLS) data, and Bank of Canada public records. No private API keys are stored in the repository, ensuring that user security and data privacy remain paramount during local deployment.