Augmenting GDPNow: A Modular Architecture for Foundation-Model, LLM-Augmented, and Alternative-Data-Enriched GDP Nowcasting | Brian Curry

Working Paper — May 2026

Abstract

The Atlanta Fed’s GDPNow model is the most-watched real-time estimate of U.S. GDP growth. Its architecture, fully documented in Higgins (2014) and updated regularly since, combines bridge equations, a linear single-factor dynamic factor model, Bayesian vector autoregressions, and Fisher chain-weighted aggregation in a bottom-up tracker over thirteen GDP subcomponents. The model is transparent, well-calibrated against the BEA’s advance estimate, and reliably outperforms naive statistical benchmarks — but it has not benefited from the advances of the past five years in deep representation learning, time-series foundation models, large language models, or alternative data. This paper proposes a modular augmentation architecture that preserves GDPNow’s bottom-up structure and its faithful mimicry of BEA methodology while addressing five specific limitations: (i) the linearity of the dynamic factor model, (ii) the absence of unstructured-text signals, (iii) the inability to adapt within a quarter, (iv) chronic weakness in inventory and net-exports forecasting, and (v) the lack of calibrated probabilistic output. We describe seven modules — a deep dynamic factor model, a foundation-model layer for filling missing monthly source data, an LLM signal extraction pipeline, an alternative-data integration layer, a probabilistic combination and online-learning layer, an inventory subsystem overhaul, and a revision-backcasting module — and outline an integrated system with a clear evaluation methodology. We discuss data-leakage risks specific to time-series foundation models and large language models, and propose evaluation protocols that respect real-time release calendars.

Keywords: nowcasting, mixed-frequency forecasting, dynamic factor models, foundation models, large language models, alternative data, GDP

JEL: C53, C55, E37

1. Introduction

GDPNow has, since 2011, served as the canonical real-time estimate of U.S. real GDP growth ahead of the BEA’s advance release. Its design — eight pipeline steps over thirteen GDP subcomponents, glued together with bridge equations, BVARs, a dynamic factor model, and Fisher chain-weighting — is itself a careful synthesis of two econometric traditions: the bridge-equation tradition of Klein, Miller-Chin, and Macroeconomic Advisers, and the dynamic factor tradition of Stock-Watson, Doz-Giannone-Reichlin, and Giannone-Reichlin-Small. Higgins (2014) gives the canonical specification; the Atlanta Fed’s “Modifications” document records its evolution through the COVID period, the 2025 gold-imports anomaly, the September 2025 NIPA data-center disaggregation, and the October-November 2025 government shutdown imputation.

What GDPNow does well is also well-understood: it provides a high-frequency, no-judgment, fully-documented nowcast that mimics the BEA’s data-construction logic faithfully enough to produce 0.77 percentage-point average absolute error and 1.17 percentage-point RMSE against the advance estimate across 2011:Q3 through 2025:Q2. What it does not yet do — and what is the subject of this paper — is incorporate any of the methodological advances of the last five years. The dynamic factor model is single-factor and linear. The forecasting equations are factor-augmented autoregressions with AIC-selected lags. The only “unstructured” data the model touches are processed survey diffusion indices (ISM, regional Fed manufacturing surveys); the model has no exposure to firm-level commentary, news flow, or the Beige Book itself. Forecasts are point estimates, not predictive distributions. Combination weights between bridge and BVAR forecasts are weighted-least-squares estimates, fixed once a quarter begins.

This paper proposes a modular augmentation architecture for GDPNow that uses modern machine learning where it adds value and leaves the existing pipeline alone where it does not. The contributions are:

A specification of where exactly classical methods can be improved by modern alternatives — deep representation learning for the factor model, time-series foundation models for univariate forecasting of monthly source data, LLM-derived signals from unstructured text, and probabilistic combination with online adaptation.
A modular architecture that preserves the BEA-mimicking bottom-up structure (Fisher chain-weighting, the 13-subcomponent decomposition, the bridge equation framework) and treats each ML/AI augmentation as a drop-in replacement for a specific upstream signal, not a wholesale replacement of the model.
A discussion of evaluation methodology that addresses two failure modes specific to modern methods in macro forecasting: data leakage in time-series foundation model benchmarks (Rambachan 2025; Auer et al. 2025), and training-cutoff contamination in LLM-derived signals.
An honest accounting of where modern ML methods have demonstrated gains for macro nowcasting — Beige Book sentiment indices have produced significant recession-forecasting accuracy (Filippou et al. 2024; Gascon and Martorana 2024); LLM forecasts of conditional inflation paths are competitive with professional forecasters (Faria-e-Castro and Leibovici 2024); time-series foundation models are competitive with but rarely dominate well-tuned classical models on benchmark suites — and where they remain speculative.

The paper proceeds as follows. Section 2 reviews the GDPNow baseline and identifies five concrete limitations. Section 3 lays out the design principles for any augmentation. Sections 4–10 describe each of the seven proposed modules. Section 11 describes the integrated architecture. Sections 12–14 discuss evaluation, implementation, and limitations. Section 15 concludes.

2. Baseline: GDPNow Architecture and Limitations

2.1 The eight-step pipeline (in brief)

GDPNow forecasts GDP by aggregating 13 subcomponents using BEA’s Fisher chain-weighting. Step 1 fits a 5-lag quarterly BVAR over the 13 real-quantity components and a parallel BVAR over their deflators, with Minnesota priors following Bańbura, Giannone, and Reichlin (2008). Step 2a fits a 12-lag 34-variable price BVAR with Waggoner-Zha (1999) conditional forecasting for staggered releases. Step 2b fits a single-factor dynamic factor model over ~125 monthly series via two-step Doz-Giannone-Reichlin estimation plus Kalman smoothing, with an idiosyncratic AR(1) added in April 2018 to dampen ISM-release jumps. Step 3 uses the smoothed factor in factor-augmented autoregressions to fill missing monthly observations. Step 4 runs bridge-equation regressions on each granular subcomponent of the five investment and government categories, aggregates via Törnqvist weights, and combines the bridge forecast with the Step-1 BVAR via restricted weighted least squares with exponential observation decay (after the January 2023 modification). Step 5 handles net exports — including a bespoke gold-imports adjustment added in March-April 2025. Step 6 handles consumption via the retail-control mapping and a CPI-to-PCE crosswalk. Step 7 handles inventories through a coupled system of price BVARs, block BVARs, and a maximum-likelihood IVA-estimation procedure that combines book-value identities with PPI-based BEA methodology. Step 8 aggregates with Fisher chain-weighting.

2.2 Five limitations

L1: Linear, single-factor representation of the monthly economy. The dynamic factor model is single-factor, linear-Gaussian, and treats heterogeneity in series via independent idiosyncratic AR(1) processes. Multi-factor extensions and nonlinear factor models have been shown to improve nowcasting performance in related contexts (Bańbura and Modugno 2014; Marcellino and Sivec 2021); deep state-space models and structured state-space sequence models (Gu, Goel, and Ré 2022) offer richer dynamics with tractable inference.

L2: No use of unstructured text data. The model touches no qualitative information. The Beige Book, FOMC statements, sectoral earnings call transcripts, and high-frequency news flow are all real-time information sources that contain forward-looking signals about the same economic activity GDPNow is trying to measure. The applied literature has demonstrated that LLM-derived sentiment from Beige Book text alone produces recession-classifier F1 scores up to 0.89 at contemporaneous horizons (Filippou et al. 2024).

L3: Frozen during the quarter. The current model freezes both parameters and methodology once a quarter begins, by design — the Atlanta Fed describes this as a discipline to prevent outcome-driven retuning. The trade-off is that the model cannot adapt to structural breaks in real time. COVID-19 required multiple methodology changes that were rolled out only between quarters, in retrospect. A principled online adaptation mechanism — one that responds to detected regime change without producing the kind of overfitting the Atlanta Fed is trying to avoid — would be valuable.

L4: Chronic weakness in inventory and net-exports forecasts. Higgins’s own forecast-error decomposition (Tables 1–6 of the working paper) shows that inventories and net exports have the largest contribution variance and the largest contribution to forecast-error variance across every forecaster examined, including the Federal Reserve Board’s Greenbook. These two components are also the ones GDPNow forecasts most indirectly — through the IVA likelihood maximization and through monthly-data factor-augmented ARs on net-exports contributions. There is significant room for improvement using firm-level data, freight and shipping data, and sector-specific time-series models.

L5: Point forecasts, no calibrated uncertainty. GDPNow publishes a single number plus its evolution over time. There is no published predictive distribution. For policymakers and risk managers, calibrated quantile forecasts (P10, P25, P50, P75, P90) would convey meaningfully more information at modest cost, particularly during regime transitions. Conformal prediction methods (Stankeviciute, Alaa, and van der Schaar 2021; Xu and Xie 2023) and probabilistic foundation models (Ansari et al. 2024 for Chronos) make this directly tractable.

3. Design Principles

Before specifying modules, we set five design principles that any augmentation should respect.

P1: Preserve BEA mimicry. GDPNow’s core value proposition is that it constructs a number using the same accounting identity and data sources the BEA will use 25–30 days later. The bottom-up structure, the 13-subcomponent decomposition, the Fisher chain-weighting, and the use of identical source data must be preserved. Augmentations enter as improved signal-generation processes upstream of the aggregation step, not as black-box replacements of it.

P2: Modular augmentation, not replacement. Each augmentation module should produce an output drop-in compatible with one specific upstream signal in the existing pipeline. A deep DFM replaces only the Step-2b factor; a TSFM replaces only the Step-3 monthly imputation; LLM signals enter the bridge equations as additional regressors. This keeps the augmented system auditable and rollback-friendly.

P3: Honor the real-time release calendar. Every signal must have a clear release timestamp and a documented vintage. LLM-derived signals must be constructed from text available no later than the corresponding model run. This rules out using LLMs whose training data crosses the relevant horizon (Rambachan 2025); fixed-cutoff or local LLMs are required for historical evaluation.

P4: Explainability constraint. Because GDPNow is a public communication instrument and is consumed by FOMC participants, components must be auditable post-hoc. A 70 billion-parameter transformer that produces a quarter-changing forecast contribution is acceptable only if the contribution can be decomposed into named inputs (e.g., “the freight tonnage signal contributed +0.2pp to the Q3 nowcast”). Attribution methods (integrated gradients for neural state-space; SHAP for tree-based methods; attention-weight inspection for transformers) should be standard outputs.

P5: Reproducibility and vintage handling. ALFRED-style vintage data must drive every backtest. Real-time backtests with appropriately vintaged data have been a stumbling block for the academic literature; they are not optional here.

4. Module A: Deep Dynamic Factor Model

4.1 Limitations of the linear DFM

The Step-2b DFM in GDPNow takes the form:

State dynamics:   f_t = rho_1*f_{t-1} + rho_2*f_{t-2} + rho_3*f_{t-3} + u_t
                  u_t ~ Normal(0, sigma_u^2)

Measurement:      z_i_t = gamma_i * f_t + eps_i_t
                  eps_i_t = phi_i * eps_i_{t-1} + eta_i_t       (idiosyncratic AR(1))

where:
  f_t       = scalar latent factor at month t
  z_i_t     = standardized value of series i at month t (~125 series)
  rho_k     = factor AR coefficients (k = 1, 2, 3)
  gamma_i   = factor loading for series i
  phi_i     = idiosyncratic AR(1) coefficient for series i (added April 2018)

The factor is estimated jointly from the ~125 standardized monthly series via two-step Doz-Giannone-Reichlin plus Kalman smoothing. The model imposes linearity in both the state dynamics and the observation equation. The single-factor assumption discards systematic variation that is not perfectly correlated with the dominant principal component. Nonlinear comovement — for example, the asymmetric response of services and goods to a supply shock, or the regime-dependent loading of manufacturing series on the common factor — is structurally invisible.

4.2 Proposed: nonlinear state-space model with structured state-space dynamics

Replace the linear factor dynamics with a small-capacity nonlinear state-space model. The state vector f_t (dimension k = 3 to 8, chosen by predictive likelihood) evolves under:

State dynamics:   f_t = g_theta(f_{t-1}, f_{t-2}, ...) + u_t

Measurement:      z_i_t = h_psi_i(f_t) + eps_i_t

where:
  f_t       = k-dimensional latent factor vector
  g_theta   = structured state-space layer (S4 / Mamba) or small recurrent transformer
  h_psi_i   = per-series shallow MLP mapping factor to observation
  theta, psi = learned parameters

Inference proceeds via amortized variational inference: an encoder q_phi(f_{1:T} | z_{1:T}) approximates the posterior, trained jointly with theta and psi by maximizing an evidence lower bound (ELBO).

This design preserves three properties the linear DFM had: (i) handling of jagged-edge missing data, via a missing-data-aware encoder; (ii) clean separation of factor and idiosyncratic dynamics; and (iii) Kalman-like smoothing behavior, but with the smoother now learned rather than analytical. Crucially, the encoder produces calibrated state samples, enabling downstream probabilistic propagation.

4.3 Validation against the linear DFM

The deep DFM should be evaluated against the GDPNow linear DFM on three metrics:

Reconstruction RMSE on held-out monthly series, conditional on observed series — the analog of the linear DFM’s measurement-equation residual variance.
Predictive log-likelihood of GDP subcomponents one quarter ahead.
Stability of factor estimates across vintages — a deep DFM that produces wildly different historical factors as new data arrives is unusable.

The expected improvement is modest at trend points and larger at turning points; this is the consistent pattern from the deep state-space literature.

4.4 Alternative: mixture-of-experts factor model

A second option, more conservative, is to keep the linear factor structure but route different series to different factors via a learned gating function — effectively, an MoE-DFM. This was the spirit of Moirai-MoE (Liu et al. 2024b) at the time-series level; applied at the factor level, it allows manufacturing series to load on a “goods factor”, services series on a “services factor”, and trade series on a “trade factor”, with the gating learned from data rather than imposed. This is closer to the multi-factor DFMs that have been productive in the academic literature (Stock and Watson 2002, 2011).

5. Module B: Foundation Models for Monthly Source Data

5.1 The Step-3 problem

Once the factor is estimated, GDPNow uses a factor-augmented AR to fill the missing months of each monthly source-data series:

Dlog(y_i_{t,h}) = alpha_i + sum_{k=1..q} gamma_i_k * Dlog(y_i_{t,h-k})
                          + sum_{j=0..r} beta_i_j * f_{t,h-j}

where:
  y_i_{t,h}  = source-data series i, month h of quarter t
  Dlog(.)    = first difference of natural log (~ growth rate)
  f_{t,h}    = dynamic factor from Step 2b
  q, r       = lag lengths chosen by AIC
               (q in [1, 6] after 2022 modification; r in [0, 3])

This step is a strong candidate for replacement by a time-series foundation model (TSFM).

5.2 The TSFM landscape, 2024–2026

Between late 2023 and early 2026, a cluster of large pretrained models for time-series forecasting was released. The leading current candidates:

Chronos and Chronos-Bolt (Ansari et al. 2024a, 2024b) — encoder-decoder transformer trained on tokenized scalar values; Chronos-Bolt is the inference-optimized variant.
Chronos-2 (Ansari et al. 2025) — improved coverage and calibration.
TimesFM (Das et al. 2024) — Google’s decoder-only patch-based transformer; TimesFM-2.5 is the current release. Univariate only.
Moirai and Moirai-MoE (Woo et al. 2024; Liu et al. 2024b) — Salesforce’s multivariate-capable architecture with mixture-of-experts.
Time-MoE (Shi et al. 2024) — sparse mixture-of-experts variant with strong recent benchmark performance.
MOMENT (Goswami et al. 2024) — foundation model trained for multiple downstream tasks beyond forecasting.
TiRex (Auer et al. 2025) — current state-of-the-art on several public benchmarks at time of writing.

Across published benchmarks, Chronos-Bolt and Time-MoE consistently emerge as Pareto-frontier on accuracy-vs-latency. Crucially, none of these models dominate well-tuned classical methods uniformly: on the German electricity-price benchmark, for example, a simple biseasonal MSTL still outperforms every TSFM tested. On macroeconomic series in particular, the consensus is that TSFMs are competitive but not dominant in zero-shot mode; fine-tuning typically helps.

5.3 Proposed integration

We propose to replace Step 3’s factor-augmented AR with a hybrid:

y_hat_i_{t,h} = lambda * TSFM(y_i history, context)
              + (1 - lambda) * FAR(y_i, f)

where:
  lambda     = per-series weight, learned (allowed to be zero,
               in which case the existing FAR is preserved)
  TSFM(.)    = foundation model forecast given series history
  FAR(.)     = the existing factor-augmented AR from Step 3

The TSFM contribution is the foundation model’s forecast conditional on the available history of series i; the context window includes recent lagged values and, for multivariate-capable models, related series.

Two important caveats:

Caveat 1: Multivariate limitation. Most current TSFMs (TimesFM, Chronos, Lag-Llama) are univariate. They do not natively handle the cross-series dependencies that the factor-augmented AR captures via the shared factor f. Moirai and Time-MoE support multivariate input; these are preferred for series whose forecasts are known to be cross-dependent (e.g., manufacturing shipments and durable-goods orders).

Caveat 2: Data leakage. The TSFM benchmarking literature has documented extensive pretraining-test contamination (Auer et al. 2025; the benchmarking working group paper from October 2025). For macro series in particular, public FRED data has often been included in pretraining corpora. Any backtest must use cleanly held-out vintage data, and ideally a TSFM whose training corpus is auditable. Chronos’s training data is documented; TimesFM’s is partly proprietary. The safer choice for production use is a fine-tuned variant whose final training data is explicitly version-controlled.

5.4 Fine-tuning protocol

For each monthly source series, fine-tune the chosen TSFM on the full vintage-respecting history with a sliding-window protocol. Cross-validation must use a forward-chaining split. Hyperparameter selection should optimize directly on the downstream metric — the GDP nowcast error contribution of the series, not the series-level forecast error — because the two are not perfectly aligned.

6. Module C: LLM-Derived Signals

This module is where the largest information gain is plausibly available, because GDPNow currently uses zero unstructured-text input. Three signal types are most promising.

6.1 Beige Book sentiment

The Beige Book is published eight times per year, two weeks before each FOMC meeting, and reports qualitative regional economic conditions across all 12 Federal Reserve Districts. Recent work has established that LLM-extracted sentiment from the Beige Book is a useful real-time recession indicator: Filippou et al. (2024) at the Cleveland Fed construct district-level and topic-level sentiment indices that exhibit F1 scores up to 0.89 for contemporaneous NBER-recession classification. Gascon and Martorana (2024) at the St. Louis Fed construct recession-probability indices from Beige Book anecdotes. The Boston Fed’s November 2025 piece extends this to FinBERT-Tone over the full 1970–2025 history.

For GDPNow, the relevant integration is more granular than recession classification. Topic-level Beige Book sentiment — specifically in the “Manufacturing”, “Construction and Real Estate”, “Retail and Consumer Spending”, “Transportation”, and “Wages and Prices” sections — can be mapped to corresponding GDP subcomponents. Concretely:

Beige Book topic Mapped GDP subcomponent Manufacturing Real PCE goods, equipment investment, IPP investment Construction and real estate Residential investment, nonresidential structures Retail and consumer spending Real PCE goods, real PCE services Transportation and logistics Real net exports (goods), inventories Wages, prices Implicit deflators (via the price BVAR in Step 2a) Banking and finance Aggregate factor (Module A); equipment investment

For each subcomponent, the topic-level sentiment index s_c_t enters as an additional regressor in the bridge equation:

Dlog(Y_c_t) = phi_c + sum_k theta_c_k * Dlog(Y_hat_c_k_t)
                    + mu_c * s_c_t + eps_t

where:
  Y_c_t        = quarterly real growth of GDP subcomponent c
  Y_hat_c_k_t  = forecast of the kth monthly indicator for subcomponent c
  s_c_t        = topic-level Beige Book sentiment for subcomponent c
  mu_c         = sentiment loading, shrunk toward zero by a ridge prior

The ridge prior on mu_c allows the data to determine the weight assigned to the qualitative signal. The sentiment series should be constructed using a single, fixed, locally-hosted model with an explicit knowledge cutoff to avoid look-ahead contamination.

6.2 FOMC and Fed staff text

The FOMC statement, the FOMC minutes (released three weeks after each meeting), and Fed speeches are all relevant signals. Hansen, McMahon, and Prat (2018) and the broader literature on Fed text have established that minute-level sentiment moves markets and forecasts policy actions. For GDP nowcasting specifically, the FOMC’s own communication contains information about staff projections of the current quarter; extracting this requires LLM-driven structured extraction rather than sentiment scoring.

A specific signal worth constructing: a “Fed assessment vector” extracted from each FOMC statement and minute, scoring along each of the 13 GDP subcomponents (consumption strength, residential investment, business investment, government spending, net exports, inventories). This is a structured-extraction task — given the text, return a JSON object with scores. Modern instruction-tuned LLMs do this reliably with calibrated few-shot prompting.

6.3 Earnings call commentary

Public-company earnings calls are released approximately one month after quarter end and contain sector-specific forward-looking commentary on demand, inventories, capital expenditure plans, and pricing. For Q1 GDP nowcasting, Q4 earnings calls (released January through early March) are real-time signals; Q1 earnings calls (released late April onward) are not.

The integration is:

Identify sector mappings (S&P 500 sectors → GDP subcomponents — e.g., consumer discretionary and consumer staples → PCE goods/services; industrials → equipment investment, IPP investment, freight; materials → inventories and net exports).
For each call, run a structured-extraction prompt asking for forward-looking statements about demand, inventory levels, capital expenditure plans, and pricing, returning a numeric score per category.
Aggregate scores within each sector with revenue-weighted averages.
The resulting sector-level signal vector enters the corresponding bridge equation as an additional regressor (or as a regressor on the contribution-to-growth specification used for net exports and inventories).

Pham and Cunningham (2024) and Carriero, Pettenuzzo, and Shekhar (2025) demonstrate that narrative prompting of LLMs produces useful forecasts for inflation and unemployment; the same machinery extends to subcomponent forecasting if the prompts are carefully designed and the outputs are calibrated.

6.4 News event extraction

The fourth signal type is real-time event detection from news flow. Supply chain disruptions, weather events, port strikes, geopolitical shocks — all of these affect specific GDP subcomponents on specific horizons, and all are detectable from news flow in advance of the corresponding statistical release.

Architecture: an LLM agent ingests a curated news feed (Reuters, Bloomberg, AP, regional papers), runs structured extraction to identify economic-relevance events tagged with (category, magnitude, geography, expected duration, affected GDP subcomponent), and produces a daily event vector. The event vector enters the model as a contemporaneous regressor on the affected subcomponent.

This is the riskiest of the three LLM signal types, because event extraction is noisy and the historical signal-to-noise ratio is hard to establish without extensive evaluation. The mitigation is to constrain the model to a narrow event taxonomy (no more than 20 distinct event categories) and to require multi-source corroboration before an event is added to the signal vector.

6.5 Look-ahead and contamination

The single largest evaluation risk in this module is look-ahead bias from LLMs whose training data spans the historical evaluation period. Rambachan (2025) proposes the use of open-source LLMs with documented training cutoffs as the standard solution; we adopt this recommendation. Specifically, all historical sentiment indices used in evaluation must be constructed from an LLM whose training cutoff predates the corresponding evaluation period, OR from a classical method (FinBERT, Loughran-McDonald) with no training-data leakage concern. Modern instruction-tuned LLMs are appropriate only for prospective (out-of-sample, forward-only) signal construction.

7. Module D: Alternative Data Pipeline

GDPNow uses only data that the BEA itself uses or could use. There is a substantial set of high-frequency private and semi-public data sources that the BEA does not yet use but that are available to private nowcasters and have demonstrated nowcasting value.

7.1 High-frequency spending

Credit and debit card data from Visa, Mastercard, Affirm, and bank-issued card programs (BAC, JPM) provides daily-to-weekly spending data that leads the Census retail sales report by approximately three weeks. The Opportunity Insights real-time spending tracker (Chetty et al.) demonstrated the value of this data during COVID-19. For GDPNow, weekly card-spending growth — decomposed by category (essentials, non-essential goods, services, travel, healthcare) — maps directly to PCE subcomponents in Step 6. The signal enters as a leading indicator in the PCE forecasting equation.

7.2 Freight and shipping

Container freight rates (Drewry WCI, Freightos FBX), truck tonnage (ATA index), and rail carloads (AAR) are weekly real-time indicators of physical economic activity. Each maps to specific GDP subcomponents:

Source Subcomponent Container freight (FBX, WCI) Real goods exports/imports Truck tonnage Real PCE goods, inventories Rail carloads (intermodal, coal, grain) Inventories, net exports Port activity (LA, Long Beach, NY/NJ) Real goods imports

These should enter the existing bridge equations as additional regressors. They are particularly useful for net exports, which is consistently the worst-forecast subcomponent.

7.3 Labor market high-frequency data

Indeed and LinkedIn job postings, ADP private payrolls, and Homebase shift data are real-time labor-market indicators that lead the Census Household Survey and the BLS Employment Situation report. The Atlanta Fed’s own Wage Growth Tracker uses CPS micro-data with a lag; high-frequency wage data from payroll providers (Gusto, ADP) would supplement this.

7.4 Mobility and activity

Google’s COVID-era Mobility Reports were discontinued in 2022, but commercial replacements exist (Cuebiq, Foursquare/Placer.ai, SafeGraph). Foot-traffic data to retail establishments leads PCE goods by approximately two weeks. Air-travel volumes (TSA daily throughput) lead PCE services (travel) by zero days — they are essentially real-time.

7.5 Satellite-derived signals

Three satellite-derived signals are worth incorporating: (i) nightlights radiance (proxy for industrial production); (ii) port-vessel activity (proxy for goods trade flows; available from Marine Traffic, Spire, AIS feeds); (iii) thermal imaging of refineries and industrial facilities (proxy for utilization rates; available from commercial providers). Each is a weak signal on its own but adds incrementally when combined with the existing factor.

7.6 LLM-orchestrated ingestion

A single LLM agent orchestrates the ingestion of these heterogeneous sources. Its responsibilities are: (i) fetching and parsing each release on its native cadence; (ii) handling format changes and source disruptions (e.g., when Marine Traffic changes its API); (iii) flagging anomalies for human review; (iv) producing structured outputs that drop into the model’s signal layer. The LLM is not making any modeling decisions — it is acting as a structured-extraction and orchestration layer over otherwise heterogeneous data sources.

8. Module E: Probabilistic Combination and Online Learning

8.1 Replacing restricted weighted least squares

The current model combines bridge and BVAR forecasts via restricted weighted least squares with exponential observation decay (the January 2023 modification). This produces a single point forecast per subcomponent. The proposed replacement is Bayesian model averaging over a larger set of candidate models — bridge, BVAR, TSFM, LLM-augmented bridge — with weights determined by recent predictive performance, optionally with regime-dependent gating.

Specifically, let Y_hat_c_m_t denote the forecast of subcomponent c from candidate model m at time t. The combined forecast and its weight update are:

Combined forecast:    Y_hat_c_t = sum_m  w_c_m_t * Y_hat_c_m_t

Weight update:        w_c_m_t  proportional to
                          w_c_m_{t-1} * p(actual_c_{t-1} | Y_hat_c_m_{t-1}, sigma_m)

where:
  Y_hat_c_m_t  = forecast of subcomponent c from model m at time t
  w_c_m_t      = posterior weight on model m at time t (sums to 1 across m)
  p(. | .)     = Gaussian predictive density evaluated at the realized value
  sigma_m      = predictive standard deviation of model m

This is the standard Bayesian model averaging recursion (Raftery, Madigan, and Hoeting 1997; in the macro forecasting context, Koop and Korobilis 2012). It produces both a point forecast and a predictive distribution that incorporates between-model uncertainty.

8.2 Online changepoint detection

To address Limitation L3 (frozen during the quarter), we propose adding online Bayesian changepoint detection (Adams and MacKay 2007) over the residuals of the combined model. When a changepoint is detected — that is, when the posterior probability of a regime shift exceeds a threshold — the weights w_c_m_t are reset to a less informative prior, and recent observations are upweighted. This produces an adaptive system that responds to structural breaks without manual intervention, while still retaining the “no judgmental adjustment” discipline (the changepoint detector is itself a published algorithm with no hidden tuning).

8.3 Conformal prediction for marginal coverage

For each subcomponent forecast, conformal prediction (Vovk, Gammerman, and Shafer 2005; Stankeviciute et al. 2021) produces calibrated prediction intervals with finite-sample coverage guarantees, conditional only on exchangeability. For time-series, the appropriate variant is conformal prediction with adaptive coverage (Gibbs and Candès 2021; Xu and Xie 2023), which maintains nominal coverage even under distribution shift. The output is a calibrated P10-P90 interval per subcomponent, which aggregates to a calibrated interval on GDP growth via Monte Carlo through the Fisher chain-weighting layer.

8.4 Mixture-of-experts gating

A more aggressive version of online adaptation uses a small neural gating function (a 2–3 layer MLP) to learn time-varying weights. The gate takes as input recent forecasting performance, the dynamic factor state, and detected changepoint signals, and outputs a soft weighting over candidate models. This requires more training data than the Bayesian update and risks overfitting; it should be evaluated against simpler alternatives before deployment.

9. Module F: Inventory Subsystem Overhaul

Step 7 of GDPNow — the inventory subsystem — is the model’s most complex piece and also its weakest. The forecast-error decomposition tables in Higgins (2014) consistently identify inventories as the single largest contributor to nowcast error. The current architecture combines a price BVAR, a 16-variable core quantity BVAR, six block BVARs (one per inventory industry), an IVA model based on BLS PPIs with a fitted turnover pattern and an AR(1) on residuals, and a joint maximum-likelihood estimation of IVAs given Census book-value identities.

We propose a parallel inventory subsystem using firm-level and freight-derived data, with the output combined with the existing system via Module E’s combination layer.

9.1 Firm-level inventory data

Public-company quarterly inventory disclosures (from 10-Qs) provide industry-level inventory stocks at the firm level, with approximately 30-day reporting lag. For NIPA “merchant wholesale” and “retail ex-autos” inventories, the constituent S&P 500 firms cover a substantial fraction of the total. A bottom-up estimate constructed from firm-level disclosures, weighted by revenue, can be used as a cross-check or an additional regressor on the existing IVA estimation.

Earnings call commentary on inventory levels — extracted via the Module C pipeline — provides a higher-frequency qualitative signal. Specifically, retailers’ commentary on inventory-to-sales ratios is a leading indicator of the contribution of inventories to GDP growth.

9.2 Sector-specific transformers

For each of the eight industry categories in the inventory subsystem (durable manufacturing, nondurable manufacturing, merchant wholesale, retail ex-autos, nonmerchant wholesale, retail autos, farm, construction+mining+utilities+other), fit a sector-specific transformer or Mamba state-space model that takes as input:

Industry-specific monthly source data (shipments, sales, employment, capacity utilization)
The dynamic factor from Module A
Sector-specific freight and shipping signals (Module D)
Sector-specific earnings-call signals (Module C)
Producer price indices for the sector

and produces a forecast of the monthly change in private inventories for that industry.

9.3 Combination

The new subsystem produces a parallel forecast of CIPI per industry; the legacy IVA-based system continues to operate; both feed into Module E’s combination layer with weights determined by recent predictive performance. The expected behavior is for the legacy system to dominate during stable periods (where its careful adherence to BEA methodology gives it an edge) and for the new system to dominate during turning points (where high-frequency firm and freight data lead).

10. Module G: Revision Backcasting

GDPNow explicitly does not anticipate BEA revisions to prior-quarter GDP estimates, with one exception: revisions to the change in private inventories in the prior quarter are propagated forward through the model’s running estimate of the inventory contribution. The model’s FAQ explicitly states that users should use the forecasts of the change in “net exports” and the change in the “change in private inventories”, not the levels — precisely because levels are subject to BEA revision that the model does not anticipate.

This is a significant gap. The BEA revises its GDP estimates twice in the two months after the advance release (second and third estimates), and again in the annual benchmark revision each July. The revisions are not random: they correlate with subsequent source-data releases (the M3 manufacturing report, revisions to retail sales, revisions to construction spending). A model trained on (advance estimate, current estimate) pairs, conditional on the source data available at the time of each subsequent revision, can produce useful predictions of revision direction and magnitude.

10.1 Approach

Train a sequence model — a small transformer or gradient-boosted tree ensemble — on the historical revision pair (GDP_advance_t minus GDP_current_t) as a function of:

The source-data revisions released between the advance estimate and the current estimate
Forecast residuals of the GDPNow model at the time of the advance release
The dynamic factor state at the time of each subsequent release
LLM-derived signals from FOMC commentary and Beige Book editions issued between the advance and the current estimate

The output is a predicted revision distribution per prior quarter, updated at each new data release. This is published alongside the current-quarter nowcast.

10.2 Why this matters

For policymakers, the revised history of GDP is often as important as the current-quarter nowcast. A model that publishes both a current-quarter nowcast and a calibrated estimate of how the prior two quarters’ published estimates are likely to revise — and why — would meaningfully expand GDPNow’s policy relevance.

11. Integrated Architecture

The seven modules combine into an integrated system organized into four planes.

11.1 Data plane

The data plane ingests, validates, and version-controls every input:

Structured economic releases (the existing GDPNow data ~125 monthly series + 34 price series + 13 quarterly series) — ingested from Haver/FRED with vintage tags from ALFRED.
Alternative data (Module D) — high-frequency spending, freight, mobility, satellite — ingested from commercial APIs via LLM-orchestrated agents.
Unstructured text (Module C) — Beige Book, FOMC, earnings calls, news — ingested from official feeds and licensed providers, stored with release timestamps.

Every series carries a vintage timestamp. Backtests use ALFRED-vintaged data exclusively.

11.2 Signal plane

The signal plane transforms raw data into model-ready features:

Latent factor signal (Module A — deep DFM).
Imputed monthly source data (Module B — TSFM hybrid).
LLM-derived structured signals (Module C — Beige Book sentiment, FOMC assessment vectors, earnings-call sector signals, news event vectors).
Alternative-data structured signals (Module D — high-frequency aggregates by category).

11.3 Model plane

The model plane runs the GDPNow steps 1–7 (preserved) plus the augmenting modules:

Step 1 (quarterly BVAR over 13 GDP subcomponents) — preserved.
Step 2a (price BVAR) — preserved, augmented by Module C signals on wages/prices topic.
Step 2b (DFM) — replaced by Module A.
Step 3 (factor-augmented AR for monthly imputation) — replaced by Module B.
Step 4 (bridge equations + BVAR combination) — preserved, with Module C/D signals as additional regressors, and Module E replacing the restricted-WLS combination.
Step 5 (net exports) — preserved with freight/shipping signals (Module D) and earnings-call commentary (Module C) added.
Step 6 (consumption) — preserved with card-spending data (Module D) and retail-sector earnings calls (Module C) added.
Step 7 (inventories) — preserved, with parallel Module F subsystem combined via Module E.
Module G (revision backcasting) — new, parallel to the main pipeline.

11.4 Aggregation and serving plane

Step 8 (Fisher chain-weighting) — preserved exactly. The aggregation step takes 13 subcomponent forecasts (now with predictive distributions from Module E) and produces a GDP growth distribution via Monte Carlo through the chain-weighting formula.
Output: a point nowcast (preserved for continuity with the existing public series), a calibrated predictive distribution (P10, P25, P50, P75, P90), a decomposition of the forecast change since the last update by data release and by signal source, and a revision-backcast for the prior two quarters.

11.5 Latency

The latency budget per release event:

Structured-data release ingestion and parsing: <1 minute.
Module A (deep DFM) re-smoothing: <2 minutes on a single GPU.
Module B (TSFM) forecasting across all monthly series: ❤ minutes on a single GPU (Chronos-Bolt at the recommended model size).
Module C (LLM extraction) for new text releases: 5–30 minutes depending on volume, run asynchronously.
Module D (alternative data) refresh: variable; daily for spending, weekly for freight.
Module E (combination, conformal calibration): <1 minute.
Module F (inventory subsystem): <5 minutes.
Module G (revision backcast): <1 minute.

Total wall-clock from release to updated nowcast: under 15 minutes for any single release event. This is faster than the current GDPNow update cycle (typically 2–3 hours).

12. Evaluation Methodology

The evaluation framework must address three failure modes specific to this kind of augmented model.

12.1 Vintage-respecting backtests

Every backtest must use ALFRED-vintaged data exclusively. The TSFM literature has been plagued by data leakage from pretraining corpora; the LLM literature is plagued by training-cutoff contamination. Both are avoidable with disciplined vintage handling and explicit knowledge-cutoff documentation per model.

The backtest period is 2000:Q1 through the most recent quarter with a published advance estimate, in line with Higgins (2014). The baseline is the current GDPNow specification. The augmented model is evaluated incrementally — one module at a time, then in full combination — to isolate each module’s contribution.

12.2 Metrics

Three families of metrics:

Point-forecast accuracy: RMSE, MAE, MdAE of the GDP growth nowcast against the BEA’s advance estimate. Reported separately for “current-quarter nowcast at advance release minus 1 day” and “current-quarter nowcast at 30, 60, 90 days before release”. Diebold-Mariano tests against the baseline GDPNow.
Probabilistic forecast accuracy: continuous ranked probability score (CRPS), pinball loss at P10/P50/P90, calibration (PIT histograms), and the interval score for the 80% predictive interval. These are reported as marginal improvements over the baseline’s bootstrap-based intervals (which would be the natural counterpart for the current point-only GDPNow).
Component-level decomposition following Higgins’s framework: the GDP RMSE is decomposed into contributions from each subcomponent forecast and their cross-covariances. This identifies which subcomponents the augmentation actually improves; “St. Offset” can mask underlying improvements if the headline is the only metric examined.

12.3 Test for genuine information gain

Each LLM-derived signal must be tested for genuine ex-ante information content, not just in-sample correlation with GDP. The protocol is:

Construct the signal using only an LLM whose knowledge cutoff predates the evaluation period.
Run a Diebold-Mariano test of the augmented bridge equation against the unaugmented bridge equation for the corresponding subcomponent.
Run a Hansen-McMahon-style placebo test: shuffle the text inputs across dates and re-evaluate. Genuine information gain should disappear under shuffling.

The same protocol applies to TSFM-based augmentations: a TSFM showing in-sample improvement might be exploiting pretraining-data leakage; a shuffle test or a hold-out using post-pretraining-cutoff data is required.

13. Implementation Considerations

13.1 Compute

The augmented model requires modest additional compute relative to the baseline:

One A100 or H100 GPU is sufficient for Module A, B, and F inference.
LLM inference for Module C is the largest variable cost: full processing of all relevant text releases requires approximately $5–50 per update cycle depending on LLM choice and prompt complexity. Locally-hosted models (Llama 3 70B class) reduce this to electricity cost.
Module G adds negligible compute.

Total infrastructure: one GPU server (A100 or H100), one orchestrator (CPU), one database (Postgres + vintage-tagged time series). This is well within the operational envelope of a Federal Reserve Bank research department.

13.2 Vintage handling

ALFRED-style vintage handling is required throughout. Every signal must be reconstructible from data available at the time of the model run. This is engineering work, not research — it requires (i) a vintage-aware data store; (ii) a deterministic build pipeline; (iii) a snapshot mechanism that captures the full input state at each update.

13.3 Open-source feasibility

The augmented model is open-sourceable in a way the current GDPNow is not. The current model’s “no code released” policy is partly historical (the codebase is bespoke Atlanta Fed infrastructure) and partly justified (Haver Analytics data licenses prevent redistribution). The augmented model can be released as:

Open-source code, with Haver-specific data ingestion replaced by FRED-only ingestion (acknowledging some loss of coverage).
Open-source LLM components for Module C (Llama or equivalent), avoiding API dependence.
Documentation of every alternative-data source with public alternatives where available.

A reference implementation would not match the production model exactly because of data-license differences, but would be sufficient for replication of every methodological claim.

14. Risks and Limitations

14.1 Explainability for policy use

GDPNow is consumed by FOMC participants and policy economists who need to understand why the nowcast moved. A model that produces a 50-basis-point swing on the back of a transformer attention weight is unacceptable for policy use unless the swing can be attributed to specific named inputs. This is why Module C and Module D both include explicit attribution: every signal that enters the model has a name, a release timestamp, and an attribution weight at every update.

14.2 LLM hallucination

LLM-based event extraction (Module C, especially Section 6.4) is the weakest link. LLMs hallucinate. Mitigation: (i) require multi-source corroboration for events; (ii) constrain to a narrow event taxonomy; (iii) periodic human audit of extracted events; (iv) the signal is a small additive contribution, not a primary forecast input.

14.3 TSFM data leakage

Discussed above (Section 5.3). Mitigation: prefer TSFMs with documented training corpora, fine-tune on vintaged data, shuffle-test for leakage signals.

14.4 Sample size

Macroeconomic forecasting has a fundamental sample-size problem: the U.S. has fewer than 80 years of quarterly GDP data, fewer than 30 years of high-quality monthly source data for many series, and fewer than 15 years of useful alternative data. Deep models with millions of parameters cannot be trained from scratch on this — they must be fine-tuned from pretraining on related (or simulated) data. The risk of overfitting is real and must be controlled by aggressive regularization, early stopping, and out-of-sample evaluation discipline.

14.5 Vendor dependence

Modules C and D introduce dependencies on commercial data providers (credit card aggregators, freight data, satellite). Vendor changes — pricing, methodology, discontinuation — can disrupt the model. Mitigation: maintain at least two providers per signal category where feasible; design fallback paths that gracefully degrade to the unaugmented model when alternative data is unavailable.

14.6 The “no judgmental adjustment” discipline

The current GDPNow’s discipline of no judgmental adjustment is part of its credibility. Augmentation by ML methods is not the same as judgmental adjustment — every component is a fixed algorithm — but the perception risk is real. The Atlanta Fed (or whoever operationalizes such a model) must commit publicly to the same discipline: methodology is fixed when a quarter begins; updates between quarters are documented; no outcome-driven retuning.

15. Conclusion

GDPNow is a remarkable piece of econometric infrastructure: bottom-up, transparent, BEA-mimicking, and stable for fifteen years. Its weaknesses are well-identified — linearity in the factor model, no use of unstructured text, no online adaptation, chronic inventory and net-exports error, point forecasts only. None of these are unfixable; all of them are addressable with the methods of the past five years.

The architecture proposed here is intentionally conservative. It does not replace any part of the existing pipeline that works. It adds modules that augment specific upstream signals — the dynamic factor, the imputation of missing monthly data, the bridge equation regressors, the combination layer, the inventory subsystem — and it preserves the Fisher chain-weighting aggregation that gives GDPNow its faithfulness to BEA methodology.

The most likely outcome of full implementation, based on the magnitudes of demonstrated improvements in adjacent literatures, is a modest reduction in RMSE (perhaps 10–20 basis points on a baseline of 115 basis points), a larger improvement at turning points and during structural breaks, calibrated probabilistic forecasts where none currently exist, anticipated revisions to prior-quarter estimates, and a 5–10x reduction in update latency. None of these individually justifies the engineering effort; their combination plausibly does.

The deeper contribution, however, is methodological: a working demonstration that modern ML methods can augment a long-standing, well-specified, transparent econometric model without overturning it. The current debate in macro forecasting tends to be framed as classical-versus-ML. The architecture here is neither. It is classical at the aggregation layer, modern at the signal layer, and Bayesian at the combination layer. Each layer does what it does best. That is the right way to combine these traditions.

About the Author

Brian Curry is a Kansas City-based Data Scientist, AI expert, and founder of Vector1 Research, where he explores the frontier of Artificial Intelligence, Causal Inference, Agentic Systems, and Economic Modeling.

His work focuses on building intelligent systems that reveal the causal structures driving performance across marketing, content, and economics. He is the creator of PyCausalSim (causal discovery through simulation), Papilon (complex system optimization), Daedalus (causal and economic analysis), and Memory-Node Encapsulation (data structures for artificial episodic memory).

References

Adams, R. P., and MacKay, D. J. C. (2007). “Bayesian Online Changepoint Detection.” arXiv:0710.3742.

Ansari, A. F., Stella, L., Turkmen, C., Zhang, X., Mercado, P., Shen, H., Shchur, O., et al. (2024a). “Chronos: Learning the Language of Time Series.” arXiv:2403.07815.

Ansari, A. F., Turkmen, C., Shchur, O., and Stella, L. (2024b). “Fast and Accurate Zero-Shot Forecasting with Chronos-Bolt.” Amazon Science.

Ansari, A. F., et al. (2025). “Chronos-2.” Amazon Science.

Auer, A., et al. (2025). “TiRex: A Foundation Model for Time Series.” arXiv preprint.

Bańbura, M., Giannone, D., and Reichlin, L. (2008). “Large Bayesian VARs.” ECB Working Paper 966.

Bańbura, M., and Modugno, M. (2014). “Maximum Likelihood Estimation of Factor Models on Datasets with Arbitrary Pattern of Missing Data.” Journal of Applied Econometrics 29(1), 133–160.

Boston Fed (2025). “The Beige Book’s Value for Forecasting Recessions.” Current Policy Perspectives, November.

Carriero, A., Clark, T. E., and Marcellino, M. (2012). “Real-Time Nowcasting with a Bayesian Mixed Frequency Model with Stochastic Volatility.” Federal Reserve Bank of Cleveland Working Paper 12–27.

Carriero, A., Pettenuzzo, D., and Shekhar, S. (2025). “Macroeconomic Forecasting with Large Language Models.” Working paper, presented at the 2025 NBER Summer Institute.

Cascaldi-Garcia, D., Luciani, M., and Modugno, M. (2024). “Lessons from Nowcasting.” In Macroeconomic Forecasting, Edward Elgar Publishing.

Chen, J., Tang, G., Zhou, G., and Zhu, W. (2025). “ChatGPT and DeepSeek for Macroeconomic Forecasting.” Working paper.

Chetty, R., Friedman, J. N., Stepner, M., and the Opportunity Insights Team (2024). “The Economic Impacts of COVID-19: Evidence from a New Public Database Built Using Private Sector Data.” Quarterly Journal of Economics, forthcoming.

Das, A., Kong, W., Sen, R., and Zhou, Y. (2024). “A Decoder-Only Foundation Model for Time-Series Forecasting.” ICML 2024.

Doz, C., Giannone, D., and Reichlin, L. (2006). “A Quasi-Maximum Likelihood Approach for Large Approximate Dynamic Factor Models.” ECB Working Paper 674.

Faria-e-Castro, M., and Leibovici, F. (2024). “Artificial Intelligence and Inflation Forecasts.” Federal Reserve Bank of St. Louis Working Paper.

Filippou, I., Garciga, C., Mitchell, J., and Nguyen, M. T. (2024). “Regional Economic Sentiment: Constructing Quantitative Estimates from the Beige Book and Testing Their Ability to Forecast Recessions.” Federal Reserve Bank of Cleveland Economic Commentary 2024–08.

Gascon, C. S., and Martorana, J. (2024). “The Beige Book and the Business Cycle: Using Beige Book Anecdotes to Construct Recession Probabilities.” Federal Reserve Bank of St. Louis Working Paper 2024–037.

Giannone, D., Reichlin, L., and Small, D. (2008). “Nowcasting: The Real-Time Informational Content of Macroeconomic Data.” Journal of Monetary Economics 55(4), 665–676.

Gibbs, I., and Candès, E. (2021). “Adaptive Conformal Inference Under Distribution Shift.” NeurIPS 2021.

Goswami, M., Szafer, K., Choudhry, A., Cai, Y., Li, S., and Dubrawski, A. (2024). “MOMENT: A Family of Open Time-Series Foundation Models.” ICML 2024.

Gu, A., Goel, K., and Ré, C. (2022). “Efficiently Modeling Long Sequences with Structured State Spaces.” ICLR 2022.

Gu, A., and Dao, T. (2023). “Mamba: Linear-Time Sequence Modeling with Selective State Spaces.” arXiv:2312.00752.

Hansen, S., McMahon, M., and Prat, A. (2018). “Transparency and Deliberation Within the FOMC: A Computational Linguistics Approach.” Quarterly Journal of Economics 133(2), 801–870.

Higgins, P. (2014). “GDPNow: A Model for GDP ‘Nowcasting’.” Federal Reserve Bank of Atlanta Working Paper 2014–7.

Koop, G., and Korobilis, D. (2012). “Forecasting Inflation Using Dynamic Model Averaging.” International Economic Review 53(3), 867–886.

Liu, X., Liu, J., Woo, G., Aksu, T., Liang, Y., Zimmermann, R., Liu, C., Savarese, S., Xiong, C., and Sahoo, D. (2024b). “Moirai-MoE: Empowering Time Series Foundation Models with Sparse Mixture of Experts.” arXiv:2410.10469.

Marcellino, M., and Sivec, V. (2021). “Nowcasting GDP Growth: A Mixed Frequency Factor MIDAS Approach.” Journal of Econometrics.

Pham, T., and Cunningham, A. (2024). “Narrative Prompting for Macroeconomic Forecasting with LLMs.” Working paper.

Raftery, A. E., Madigan, D., and Hoeting, J. A. (1997). “Bayesian Model Averaging for Linear Regression Models.” Journal of the American Statistical Association 92(437), 179–191.

Rambachan, A. (2025). “Evaluating LLMs as Forecasters: Open Models, Cutoffs, and Contamination.” Working paper.

Shi, X., Wang, S., Nie, Y., Li, D., Ye, Z., Wen, Q., and Jin, M. (2024). “Time-MoE: Billion-Scale Time Series Foundation Models with Mixture of Experts.” arXiv:2409.16040.

Stankeviciute, K., Alaa, A., and van der Schaar, M. (2021). “Conformal Time-Series Forecasting.” NeurIPS 2021.

Stock, J. H., and Watson, M. W. (2002). “Macroeconomic Forecasting Using Diffusion Indexes.” Journal of Business & Economic Statistics 20(2), 147–162.

Stock, J. H., and Watson, M. W. (2011). “Dynamic Factor Models.” In The Oxford Handbook of Economic Forecasting.

Vovk, V., Gammerman, A., and Shafer, G. (2005). Algorithmic Learning in a Random World. Springer.

Waggoner, D., and Zha, T. (1999). “Conditional Forecasts in Dynamic Multivariate Models.” Review of Economics and Statistics 81(4), 639–651.

Woo, G., Liu, C., Kumar, A., Xiong, C., Savarese, S., and Sahoo, D. (2024). “Unified Training of Universal Time Series Forecasting Transformers (Moirai).” ICML 2024.

Xu, C., and Xie, Y. (2023). “Conformal Prediction for Time Series.” IEEE Transactions on Pattern Analysis and Machine Intelligence.