MACROCAST: Vintage-Consistent TSFM

• MACROCAST is a leakage-free time series foundation model that maintains vintage consistency by training solely on synthetic data and vintage-specific macro panels.
• It employs a linear-RNN architecture with state weaving and probabilistic forecasting to efficiently produce multi-horizon predictions while avoiding temporal contamination and revision bias.
• Empirical results show that MACROCAST outperforms traditional benchmarks like AR(1), underscoring its practical value for real-world macroeconomic forecasting.

Searching arXiv for MACROCAST and closely related TSFM papers to ground the article in the cited literature. First, locating the MACROCAST paper itself. MACROCAST is a lightweight, vintage-consistent time series foundation model designed specifically for real-time macroeconomic forecasting. It is defined by an end-to-end training and evaluation protocol that eliminates two leakage mechanisms that are especially consequential in macroeconomic applications: temporal contamination, in which a model has seen the realized values of the same series it later forecasts, and revision bias, in which a model is trained on fully revised data rather than the preliminary releases available to a forecaster at the time (Carriero et al., 27 Jun 2026). The model is pretrained only on purely synthetic time series and fine-tuned only on synthetic panels generated from econometric models estimated on vintage-specific ALFRED data, so no observed future or revised value enters the model at any stage (Carriero et al., 27 Jun 2026).

1. Real-time forecasting setting and leakage problem

Real-time macro forecasting is formulated around the information set actually available at the forecast date. In this setting, statistical agencies release preliminary estimates that are later revised, sometimes substantially. MACROCAST is motivated by the observation that a model trained on fully revised data has implicitly learned from information unavailable to a real-time forecaster (Carriero et al., 27 Jun 2026).

The formal setup is vintage-specific. Let $v$ denote the ALFRED vintage, interpreted as the month of release, and let the panel at vintage $v$ be $Y^{(v)}_{1:T_v} \in \mathbb{R}^{N\times T_v}$. Because revisions occur across releases, an entry $Y^{(v)}_{i,t}$ can differ across vintages $v \neq v'$. At forecast origin $t$, training uses the expanding window through $t-1$ from the vintage-$v$ panel only, and the forecast for series $i$ at horizon $h$ is

$v$0

with no observation from vintages $v$1 used in either training or evaluation (Carriero et al., 27 Jun 2026).

Within this framework, MACROCAST distinguishes two leakage types. Temporal contamination refers to training on observed values of the same target series that will later be forecast out-of-sample. Revision bias refers to training on final, revised values when only preliminary releases were available in real time (Carriero et al., 27 Jun 2026). The model’s central methodological claim is that it eliminates both forms of leakage end-to-end.

Forecast errors are evaluated against “six-month delayed actuals,” defined as the FRED-MD vintage released approximately six months after the target date. This choice is intended to avoid conflating forecast errors with subsequent revisions while also not relying on especially noisy first estimates (Carriero et al., 27 Jun 2026).

2. Architecture and forecasting mechanism

MACROCAST builds on TempoPFN’s linear recurrent neural network design rather than a transformer architecture (Carriero et al., 27 Jun 2026). In the description given for the model, this choice reduces quadratic attention costs, preserves a state-space interpretation, and enables fast per-vintage recalibration (Carriero et al., 27 Jun 2026). The recurrence is

$v$2

with input-dependent state transition matrices $v$3 and $v$4, linear in the state and nonlinear in the inputs, together with a nonlinear output map $v$5 (Carriero et al., 27 Jun 2026).

The instantiated recurrence is the GatedDeltaProduct update. The state update is stacked in three layers, and each layer is augmented with a short moving-average filter, normalization, and a nonlinear projection (Carriero et al., 27 Jun 2026). The hidden state dimension is $v$6, and the total parameter count is approximately $v$7 million (Carriero et al., 27 Jun 2026).

For multi-step prediction, MACROCAST uses state weaving, which alternates observed-history and forecast stages. In this procedure, the model feeds its own 1-step predictions back as inputs for subsequent steps, yielding multi-horizon forecasts in the econometric analogue of iterating a VAR forward (Carriero et al., 27 Jun 2026). This mechanism is paired with linear-recurrence $v$8 training complexity and single-pass multi-horizon inference (Carriero et al., 27 Jun 2026).

The model is designed for unbalanced macro panels with ragged edges. Each series is standardized, and missing observations are replaced by a learnable missing-value embedding rather than being imputed (Carriero et al., 27 Jun 2026). The training and evaluation context spans up to about ten years of monthly data, while fine-tuning operates on slices of length $v$9 months (Carriero et al., 27 Jun 2026).

MACROCAST also produces probabilistic forecasts. It predicts nine conditional quantiles per series and horizon, with $Y^{(v)}_{1:T_v} \in \mathbb{R}^{N\times T_v}$0, and uses the median $Y^{(v)}_{1:T_v} \in \mathbb{R}^{N\times T_v}$1 as the point forecast (Carriero et al., 27 Jun 2026). The training objective is the joint pinball loss over series, horizons, and quantiles:

$Y^{(v)}_{1:T_v} \in \mathbb{R}^{N\times T_v}$2

The quantiles are not post-processed to enforce monotonicity, and the reported comparisons use the median only, so quantile crossing does not affect RMSFE (Carriero et al., 27 Jun 2026).

3. Leakage-free training pipeline

MACROCAST’s training pipeline has two stages, both structured to preserve vintage consistency and exclude exposure to future or revised observations (Carriero et al., 27 Jun 2026).

The first stage is pretraining on purely synthetic time series. This stage uses approximately ten million distinct synthetic series, $Y^{(v)}_{1:T_v} \in \mathbb{R}^{N\times T_v}$3 optimization steps, and AdamW with linear warm-up and cosine decay; the reported cost is $Y^{(v)}_{1:T_v} \in \mathbb{R}^{N\times T_v}$4 hours on a single NVIDIA A6000, approximately one GPU-day (Carriero et al., 27 Jun 2026). The pretrained weights are then frozen as initialization for all fine-tuning runs (Carriero et al., 27 Jun 2026).

The pretraining generators combine the Chronos synthetic library with TempoPFN augmentations. The covered process classes include trends, seasonality, periodic oscillations, smooth kernels, regime changes, outliers, and mean-reverting dynamics (Carriero et al., 27 Jun 2026). The generator family includes trend-times-seasonality with multiplicative Weibull noise, sine-wave families, Gaussian-process kernels including SE, periodic, rational quadratic, and white noise, sawtooth and step-function processes, anomaly and spike processes, CauKer structural causal processes over a random DAG, and Ornstein–Uhlenbeck dynamics with regime-switching parameters (Carriero et al., 27 Jun 2026). Post-generation transforms include reversals, sign flips, regime insertion, amplitude modulation, calendar seasonality, resampling artifacts, TSMix convex mixing, and missing-value masks (Carriero et al., 27 Jun 2026). At each optimization step, a series and horizon from $Y^{(v)}_{1:T_v} \in \mathbb{R}^{N\times T_v}$5 to $Y^{(v)}_{1:T_v} \in \mathbb{R}^{N\times T_v}$6 steps are drawn, and the pinball loss is evaluated on held-out future points (Carriero et al., 27 Jun 2026).

The second stage is fine-tuning on synthetic panels calibrated to vintage-specific ALFRED data. For each vintage $Y^{(v)}_{1:T_v} \in \mathbb{R}^{N\times T_v}$7, the cleaned stationary panel $Y^{(v)}_{1:T_v} \in \mathbb{R}^{N\times T_v}$8 is the sole input to econometric model estimation (Carriero et al., 27 Jun 2026). MACROCAST then calibrates four generators and simulates $Y^{(v)}_{1:T_v} \in \mathbb{R}^{N\times T_v}$9-month synthetic panels with a $Y^{(v)}_{i,t}$0-step burn-in (Carriero et al., 27 Jun 2026). The four generators are a Dynamic Factor Model, a Bayesian VAR, a univariate autoregressive generator, and a block bootstrap (Carriero et al., 27 Jun 2026).

The Dynamic Factor Model is specified as

$Y^{(v)}_{i,t}$1

with idiosyncratic noise $Y^{(v)}_{i,t}$2, factor innovations $Y^{(v)}_{i,t}$3, EM estimation in state-space form via DynamicFactorMQ, factor count capped at $Y^{(v)}_{i,t}$4 with fallback to $Y^{(v)}_{i,t}$5 or $Y^{(v)}_{i,t}$6, loadings estimated by OLS, and factor dynamics estimated with VAR(2) on smoothed factors (Carriero et al., 27 Jun 2026).

The Bayesian VAR is

$Y^{(v)}_{i,t}$7

with $Y^{(v)}_{i,t}$8 and Minnesota-type priors (Carriero et al., 27 Jun 2026). Own-lag coefficients shrink toward $Y^{(v)}_{i,t}$9 for unit-root series or $v \neq v'$0 for stationary series, while cross-lags shrink toward $v \neq v'$1 with strength proportional to $v \neq v'$2. The baseline hyperparameters are $v \neq v'$3, $v \neq v'$4, and $v \neq v'$5; for the full $v \neq v'$6-variable BVAR, tighter shrinkage $v \neq v'$7, $v \neq v'$8 is used (Carriero et al., 27 Jun 2026). To accommodate early vintages with limited sample length, both clustered BVARs and a full BVAR are estimated (Carriero et al., 27 Jun 2026).

The univariate autoregressive generator estimates

$v \neq v'$9

by OLS for each series and simulates independently as

$t$0

If data are too short or coefficients are explosive $t$1, the generator falls back to small-variance white noise $t$2 (Carriero et al., 27 Jun 2026). The more general ARIMA form is also noted,

$t$3

with MACROCAST’s AR generator corresponding to $t$4 (Carriero et al., 27 Jun 2026).

The block bootstrap samples contiguous blocks from $t$5, with random starts and random lengths, concatenating them to fill $t$6 months and clipping extreme outliers greater than $t$7 (Carriero et al., 27 Jun 2026). Two regimes are used: $t$8–$t$9 months and $t-1$0–$t-1$1 months (Carriero et al., 27 Jun 2026).

Per vintage, the fine-tuning corpus contains $t-1$2 stored panels: $t-1$3 standard bootstrap, $t-1$4 long bootstrap, $t-1$5 AR, $t-1$6 DFM, $t-1$7 clustered BVAR, and $t-1$8 full BVAR (Carriero et al., 27 Jun 2026). Training examples are then constructed with stochastic augmentation by drawing a random $t-1$9-month window, selecting a random subset of $v$0 series, permuting columns, adding Gaussian noise with $v$1 per series, and forcing inclusion of outlier-prone series such as help-wanted with probability $v$2 (Carriero et al., 27 Jun 2026).

Fine-tuning uses $v$3 AdamW steps per vintage, a short linear warm-up of approximately $v$4 steps, cosine decay, gradient clipping with $v$5 norm $v$6, and mixed precision in bfloat16 (Carriero et al., 27 Jun 2026). The gradient-based fine-tuning stage completes in approximately nine minutes on a modern GPU, and the full monthly refresh including econometric estimation and simulation runs in well under $v$7 minutes (Carriero et al., 27 Jun 2026).

4. Evaluation protocol and benchmark design

The evaluation uses U.S. monthly FRED-MD data retrieved through ALFRED/FRED so that both monthly vintages and revision histories are available (Carriero et al., 27 Jun 2026). FRED-MD contains $v$8 series, of which $v$9 are excluded because they are discontinued, irregular, or have publication lags greater than two months, leaving $i$0 series (Carriero et al., 27 Jun 2026).

The evaluation window runs from August 1999 through December 2024. At each vintage month $i$1, models are trained on data through $i$2 using an expanding window (Carriero et al., 27 Jun 2026). For the ragged edge, series lagging two months are carried forward so that they have a $i$3 value, while series with lags of at least three months are excluded (Carriero et al., 27 Jun 2026). Forecasts are made for horizons $i$4 through $i$5, where $i$6 is a nowcast for month $i$7, and targets from January through June 2020 are excluded to avoid the extreme pandemic shock (Carriero et al., 27 Jun 2026).

Forecast comparison is based on RMSFE ratios relative to an AR(1) benchmark:

$i$8

with reporting at the median across series and also within category subsets (Carriero et al., 27 Jun 2026). Statistical significance is assessed with the Diebold–Mariano test using the Harvey–Leybourne–Newbold finite-sample correction, a one-sided alternative that the model beats AR(1), and Bartlett kernels with $i$9 lags for the long-run variance (Carriero et al., 27 Jun 2026). Tables use majority-vote stars when more than half of the series in a category satisfy $h$0, $h$1, or $h$2 (Carriero et al., 27 Jun 2026).

The benchmark set comprises AR(1), a BVAR with conjugate Normal–inverse-Wishart priors and hyperparameters chosen by marginal likelihood, a DFM implemented through factor-augmented direct projections with factors obtained by principal components and EM and the number of factors selected by Bai–Ng information criteria up to eight, Chronos-2 and Moirai2-Small evaluated zero-shot with public pretrained weights, and MACROCAST itself, which is pretrained once on synthetic data and fine-tuned monthly per vintage on synthetic panels estimated from that vintage’s ALFRED data (Carriero et al., 27 Jun 2026).

A notable design feature is that MACROCAST is never tuned to any single target series; it is evaluated in a series-level zero-shot mode (Carriero et al., 27 Jun 2026). This suggests that the model is intended to serve as a panel-level forecaster rather than a bank of separately specialized univariate systems.

5. Empirical results

Across the full sample, MACROCAST is reported to improve on AR(1), match or beat Chronos-2, and outperform the econometric BVAR and DFM benchmarks, while preserving the stated leakage-free guarantees (Carriero et al., 27 Jun 2026).

For the $h$3-series full-sample evaluation against six-month delayed actuals, MACROCAST’s median RMSFE ratios relative to AR(1) are $h$4 at $h$5, $h$6 at $h$7, $h$8 at $h$9, $v$00 at $v$01, $v$02 at $v$03, and $v$04 at $v$05 (Carriero et al., 27 Jun 2026). Chronos-2 records $v$06, $v$07, $v$08, $v$09, $v$10, and $v$11 at the same horizons, while BVAR and DFM are centered around one and edge above one from medium horizons onward; Moirai2-Small tracks AR(1) (Carriero et al., 27 Jun 2026).

Breadth is emphasized as a separate performance criterion. MACROCAST beats AR(1) for the majority of series at all horizons $v$12, and at $v$13 approximately $v$14 of series fall below AR(1), with a dense mass just below one and a highly compressed upper tail, compared with heavier right tails for BVAR and DFM at longer horizons (Carriero et al., 27 Jun 2026). The broader summary is that MACROCAST improves on AR(1) for roughly $v$15 of series-horizon pairs (Carriero et al., 27 Jun 2026).

On curated subsets, the model shows mixed but generally strong behavior. On the “Medium” set of $v$16 headline indicators, MACROCAST achieves $v$17 at $v$18, while Chronos-2 is best at $v$19 and $v$20 with $v$21 and $v$22 (Carriero et al., 27 Jun 2026). On the “Large” set of $v$23 indicators, MACROCAST is best at four of six horizons, including $v$24, $v$25, and $v$26 (Carriero et al., 27 Jun 2026).

The category breakdown indicates heterogeneity across macro domains. At $v$27, MACROCAST is best in Output $v$28, Labor $v$29, Money $v$30, Interest/FX $v$31, and is closest to one in Stocks; DFM is best for Consumption $v$32, and Chronos-2 leads Housing $v$33 and Prices $v$34 (Carriero et al., 27 Jun 2026). At $v$35, MACROCAST leads Output, Labor, Money, Prices, and Stocks, while Chronos-2 leads Housing and Interest/FX (Carriero et al., 27 Jun 2026). At $v$36, MACROCAST leads Output, Labor, Consumption, Interest/FX, and Stocks, while Chronos-2 remains strongest in Housing and marginally in Money (Carriero et al., 27 Jun 2026).

Several individual series are highlighted. For nonfarm payrolls (PAYEMS), MACROCAST is best across models at every horizon from $v$37 onward, with $v$38 at $v$39, $v$40 at $v$41, $v$42 at $v$43, $v$44 at $v$45, and $v$46 at $v$47 (Carriero et al., 27 Jun 2026). For industrial production (INDPRO), MACROCAST stays at or below AR(1) throughout, with examples including $v$48 at $v$49 and $v$50 at $v$51, whereas BVAR and DFM deteriorate beyond $v$52 (Carriero et al., 27 Jun 2026). For the unemployment rate (UNRATE), it is described as the most stable across horizons, with $v$53 at $v$54, $v$55 at $v$56, and approximately one at long horizons (Carriero et al., 27 Jun 2026). For the federal funds rate (FEDFUNDS), MACROCAST is strong at short and medium horizons, including $v$57 at $v$58, $v$59 at $v$60, and $v$61 at $v$62, though Chronos-2 is marginally ahead at $v$63 to $v$64 (Carriero et al., 27 Jun 2026). For capacity utilization (CUMFNS), MACROCAST is best from $v$65 to $v$66, including $v$67 at $v$68 and $v$69 at $v$70 (Carriero et al., 27 Jun 2026).

The qualitative robustness narrative is that MACROCAST’s gains are concentrated around turning points and stress episodes, especially the 2001 recession and the 2008–2009 crisis, whereas AR(1) is difficult to beat during calm expansions (Carriero et al., 27 Jun 2026). The long-horizon distributions remain anchored below one for MACROCAST, while BVAR shifts rightward, with an example median rising from $v$71 at $v$72 to $v$73 at $v$74 (Carriero et al., 27 Jun 2026).

6. Interpretation, practical use, and limitations

MACROCAST’s practical significance lies in combining leakage-free real-time training with a computational budget compatible with monthly recalibration. The pretrained model requires approximately one GPU-day once, and each vintage-specific fine-tune takes approximately nine minutes, with an end-to-end refresh time below $v$75 minutes including econometric estimation and synthetic simulation (Carriero et al., 27 Jun 2026). This makes genuine month-by-month deployment operationally feasible.

In applied use, the inputs are monthly ALFRED/FRED-MD vintages transformed to stationarity using FRED-MD transformation codes (Carriero et al., 27 Jun 2026). The ragged edge is handled by carrying forward one month for series with two-month publication lags and excluding longer-lag series (Carriero et al., 27 Jun 2026). Because missing values are handled through a learned embedding, the model supports unbalanced panels without explicit imputation (Carriero et al., 27 Jun 2026). Optimization uses AdamW with warm-up and cosine decay, gradient clipping, bfloat16 mixed precision, slices of length $v$76, and $v$77 series per batch, with small noise augmentation at $v$78 (Carriero et al., 27 Jun 2026).

The model’s central methodological contribution is not merely architectural. It is the coupling of a lightweight linear-RNN TSFM to a vintage-consistent synthetic-data regime in which pretraining uses only simulated series and fine-tuning uses only synthetic panels generated from econometric models estimated on the contemporaneous vintage (Carriero et al., 27 Jun 2026). This suggests a template for domains in which revision histories and temporal leakage are structurally important, although that broader generalization is an inference rather than a directly demonstrated result.

Several limitations are explicit. First, the macro calibration depends on synthetic panels drawn from vintage-estimated econometric models; this removes leakage and provides diverse training trajectories, but it may miss fine-grained features not captured by the parametric suite (Carriero et al., 27 Jun 2026). Second, the evaluation focuses on U.S. monthly FRED-MD data, and extension to mixed-frequency nowcasting with weekly or daily indicators is presented as a natural direction (Carriero et al., 27 Jun 2026). Third, despite regime-switching generators and bootstrap preservation of real-time co-movements, large structural breaks remain difficult; January through June 2020 are excluded for this reason (Carriero et al., 27 Jun 2026). Fourth, quantile outputs are unconstrained and may cross, so future work could impose monotone quantile layers or density calibration for improved interval forecasts (Carriero et al., 27 Jun 2026). Additional proposed extensions include structural identification, scenario analysis, expansion to other economies or domains, and hybrid fine-tuning that blends synthetic vintages with carefully curated leakage-safe observed signals such as survey nowcasts (Carriero et al., 27 Jun 2026).

7. Position within macroeconomic forecasting research

MACROCAST is situated at the intersection of foundation-model methodology and classical real-time macroeconometrics. It retains familiar macroeconometric components—Dynamic Factor Models, Bayesian VARs, autoregressive persistence, and bootstrap resampling—but uses them to generate synthetic training corpora rather than as the deployed forecasters themselves (Carriero et al., 27 Jun 2026). In that sense, the model can be viewed as a TSFM whose domain adaptation is mediated by vintage-specific econometric simulation.

Its comparative claims are correspondingly specific. Relative to Chronos-2, described as the strongest currently available TSFM in the reported comparisons, MACROCAST matches or surpasses it in aggregate and category-level evaluations while avoiding both temporal contamination and revision bias (Carriero et al., 27 Jun 2026). Relative to BVAR and DFM benchmarks, it outperforms them under the same leakage-free real-time protocol (Carriero et al., 27 Jun 2026). Relative to naive autoregressive forecasting, it improves on AR(1) for roughly $v$79 of series-horizon pairs and for approximately $v$80 of one-month-ahead forecasts (Carriero et al., 27 Jun 2026).

The broader significance of these results lies in the formulation of “vintage consistency” as a first-class design constraint for foundation models in macroeconomics (Carriero et al., 27 Jun 2026). Existing TSFMs may score well under conventional retrospective evaluation but still violate the information set available to a forecaster. MACROCAST’s contribution is to define a training, fine-tuning, and evaluation regime in which no future or revised observation is ever ingested. A plausible implication is that its empirical results should be interpreted not only as forecast-accuracy outcomes, but also as evidence that leakage-free TSFMs need not forfeit competitiveness with stronger but less strictly real-time alternatives.