MacroCast

Paper: Carriero, Pettenuzzo & Shekhar (2026), MACROCAST: A Vintage-Consistent Time Series Foundation Model for Real-Time Macroeconomic Forecasting — arXiv:2606.28670 (SSRN).

MacroCast is a compact (~1.2M-parameter) time-series foundation model for macroeconomic forecasting. It forecasts a full panel of monthly indicators in a single forward pass (no autoregressive windowing) and returns probabilistic (9-quantile) forecasts.

It is trained in two stages:

1. Base model: A ~1.2M-parameter linear-RNN backbone (built on TempoPFN, Apache-2.0), trained from scratch entirely on synthetic data (no real macro data).
2. MacroCast — Base model fine-tuned on FRED-MD via an expanding-window, real-time-vintage procedure, using a mixture of macro-calibrated synthetic panels (block-bootstrap, per-variable AR, Dynamic Factor Model, and BVAR).

Vintages (real-time fine-tuning)

This repo ships one checkpoint per fine-tuning vintage (in models/). Each checkpoint is MacroCast fine-tuned on only the FRED-MD data available through that cutoff — e.g. the 2010 vintage has never seen post-2010 data — so you can choose how much history the model has seen and run leak-free real-time / backtesting experiments. The available vintages, their data cutoffs, and per-vintage variable counts are listed in manifest.json (default_year is used when no year is given).

Usage

Install dependencies (needs a GPU). On Colab/Jupyter prefix with !; torch is already present on Colab, otherwise install the build matching your CUDA:

pip install flash-linear-attention==0.5.0 transformers==5.8.1 einops gluonts huggingface_hub

from huggingface_hub import snapshot_download
import sys

local = snapshot_download(repo_id="shubhranshu/MacroCast")
sys.path.insert(0, local)                 # for forecast.py
from forecast import MacroCastForecaster

# List vintages, then load MacroCast for one (year=None -> latest):
print(MacroCastForecaster.available_years(local))     # e.g. [1998, ..., 2023]
f = MacroCastForecaster.from_pretrained(local, year=2023)

# history: np.ndarray of shape [T, N]  (T monthly steps, N FRED-MD variables)
point, quantiles = f.predict(history, horizon=12, freq="M")

• point: [horizon, N] median (0.5-quantile) forecast.
• quantiles: [horizon, N, 9] forecasts for quantiles 0.1 … 0.9.

Model details

• Architecture: GatedDeltaProduct linear-RNN, embed=128, 3 layers, ~1.2M params.
• Objective: 9-quantile (0.1…0.9) pinball loss.
• Pre-training: synthetic priors only (GP, kernel, Ornstein–Uhlenbeck, ForecastPFN, CauKer, sine/sawtooth/step, …) — no real macro data.
• MacroCast fine-tuning: FRED-MD real-time vintages expanded into a synthetic mixture — real + block-bootstraps, arima (per-variable AR), dfm (Dynamic Factor Model), and bvar_clust / bvar_full (Minnesota BVAR).

Intended use and limitations

• Built for monthly macroeconomic panels. Feed transformed/stationary inputs.
• Forecasts are statistical outputs, not financial or policy advice.

Citation

If you use MacroCast, please cite:

@misc{carriero2026macrocast,
  title         = {{MACROCAST}: A Vintage-Consistent Time Series Foundation Model for Real-Time Macroeconomic Forecasting},
  author        = {Carriero, Andrea and Pettenuzzo, Davide and Shekhar, Shubhranshu},
  year          = {2026},
  eprint        = {2606.28670},
  archivePrefix = {arXiv},
  primaryClass  = {econ.EM},
  url           = {https://arxiv.org/abs/2606.28670}
}

Acknowledgements & license

Built on TempoPFN (Apache-2.0): the model backbone, trainer, and synthetic-data generators are vendored from that project and modified for macroeconomic forecasting. Released under Apache-2.0. Please also cite TempoPFN (Moroshan et al., 2025, arXiv 2510.25502).