An installable agent skill for macroeconomic nowcasting, central-bank policy
diagnostics, and economist-style interpretation. The installable skill lives in
economics-ml/; the dashboard and tests are project-side
supporting files.
| Layer | Main Output | Role of ML |
|---|---|---|
| GDP nowcast | Bridge-equation baseline by country and quarter | Bounded residual calibration, reported separately |
| Policy diagnostics | Base Taylor and Data-Enhanced Taylor readouts | Historical residual calibration against observable macro factors |
| Backtesting | Baseline vs calibrated OOS comparison | Prior-window calibration only, release-lag filtered |
| Agent report | Driven factors, data-through dates, validation notes | Structured interpretation, not a hidden predictor |
ML is not the primary forecaster. It is an auxiliary calibration and measurement layer used to explain where the structural model may be missing information.
Install the economics-ml/ folder for normal use. It contains SKILL.md, the
Python runtime, and requirements. The repository-level dashboard/, tests/,
and assets/ folders are not part of the installed skill.
macOS/Linux:
tmpdir="$(mktemp -d)"
git clone --depth 1 https://github.com/garroshub/ai-economist-skill.git "$tmpdir/ai-economist-skill"
rm -rf ~/.codex/skills/economics-ml
cp -R "$tmpdir/ai-economist-skill/economics-ml" ~/.codex/skills/economics-ml
rm -rf "$tmpdir"
cd ~/.codex/skills/economics-ml
pip install -r requirements.txt
Windows PowerShell:
$skillDir = "$env:USERPROFILE\.codex\skills\economics-ml"
$tmpDir = Join-Path $env:TEMP "ai-economist-skill"
Remove-Item -LiteralPath $tmpDir -Recurse -Force -ErrorAction SilentlyContinue
git clone --depth 1 https://github.com/garroshub/ai-economist-skill.git $tmpDir
Remove-Item -LiteralPath $skillDir -Recurse -Force -ErrorAction SilentlyContinue
Copy-Item -Recurse -LiteralPath "$tmpDir\economics-ml" -Destination $skillDir
Set-Location $skillDir
pip install -r requirements.txt
After installation, ask the agent for tasks such as:
Use the Economics ML skill to explain the latest US and Canada GDP nowcast drivers.
Use the Economics ML skill to compare Base Taylor and Data-Enhanced Taylor signals.
Use the Economics ML skill to audit whether the backtest has forward-information leakage.
The skill should answer with target period, data-through date, baseline estimate, data-enhanced estimate, driven factors, validation checks, and limitations. It should not merely point the user to a Python command.
If you only want the agent to produce structured interpretation from
user-provided values or existing reports, SKILL.md alone is enough. Live data
pulls and backtests require the runtime bundle above.
Clone the full repository only for development, dashboard work, or tests:
git clone https://github.com/garroshub/ai-economist-skill.git
In a full clone, dashboard/, tests/, src/, and the Python entrypoints are
project files. Only economics-ml/ is required for installing the agent skill.
-
Structural bridge model with SVD factor extraction.
-
Ragged-edge filling for missing recent monthly observations.
-
US high-frequency auxiliary variables: initial claims, housing, durable goods, real disposable income, financial conditions, and yield curve.
-
Canada auxiliary variables: StatCan retail sales, CPI, CAD/USD, WTI oil, and US demand spillover measures.
-
Baseline and ML-calibrated nowcasts shown side by side.
-
Taylor 1993, Taylor 1999, and nonlinear inflation-response variants.
-
Output-gap proxies from labor slack, HP-filtered GDP, and capacity utilization.
-
Base Taylor result remains the main structural signal.
-
Data-Enhanced Taylor result learns historical residual adjustments from activity, inflation pressure, financial conditions, external pressure, and labor cooling.
-
Enhancement decomposition is reported in percentage-point contributions.
-
Expanding-window backtest for GDP nowcasts.
-
Release-lag-aware pseudo-real-time feature filtering.
-
Baseline R2/RMSE and ML-calibrated R2/RMSE shown on the same validation window.
-
Revised-data limitation is disclosed rather than treated as a vintage-data test.
The skill can be used for analysis without running local scripts. Install the runtime dependencies only when you want live data pulls, regenerated reports, backtests, or local dashboard builds.
cd economics-ml
pip install -r requirements.txt
set FRED_API_KEY=your_key_here
Run GDP nowcasts:
python main.py gdp --country US
python main.py gdp --country Canada
Run policy diagnostics:
python main.py policy --country US
python main.py policy --country Canada
Run backtests:
python backtest_engine.py
Build the dashboard:
cd dashboard
npm ci
npm run build
- FRED API for macro indicators.
- BLS and Bank of Canada public data where available.
- Statistics Canada public CSV tables for retail-sales auxiliary features.
- Public news and outlook pages only when converted into structured measurement variables.
No private API key is stored in the repository. Set FRED_API_KEY locally before
running live workflows.
Reports should include the forecast target period, data-through date, source coverage, structural baseline, calibrated result, driven-factor decomposition, backtest window, leakage controls, and limitations. Causal claims require a separate identification design.
ai-economist-skill/
|-- README.md
|-- economics-ml/
| |-- SKILL.md
| |-- main.py
| |-- backtest_engine.py
| |-- requirements.txt
| `-- src/
|-- dashboard/
|-- assets/
`-- tests/
For skill installation, use only economics-ml/. The dashboard, tests, and
assets support the public project page and validation workflow.