arxiv: 2605.09740 · v1 · submitted 2026-05-10 · 💰 econ.EM · stat.ME· stat.ML

Recognition: 2 theorem links

· Lean Theorem

LGB+: A Macroeconomic Forecasting Road Test

Philippe Goulet Coulombe

Pith reviewed 2026-05-12 02:20 UTC · model grok-4.3

classification 💰 econ.EM stat.MEstat.ML

keywords macroeconomic forecastinggradient boostinglinear basis functionsautoregressive dynamicsforecast decompositionout-of-bag evaluationdecision treesU.S. macroeconomic data

0 comments

The pith

LGB+ improves macroeconomic forecasts by letting linear candidates compete with trees at each boosting step.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper develops LGB+, a boosting method that includes linear basis functions in the same pool as decision trees. At each iteration it either pits a linear update against a tree update and keeps only the winner according to out-of-bag performance, or follows a fixed alternating schedule of tree blocks and linear corrections. This design lets the algorithm exploit the strong autoregressive and accounting relationships common in economic series without forcing trees to approximate them through many splits. If the gains hold, forecasts decompose directly into linear and nonlinear parts and variable importance measures separate along the same lines. Readers should care because standard tree boosting wastes capacity on linear patterns in the small samples typical of quarterly macro data.

Core claim

LGB+ evaluates a tree and a linear candidate at each step against out-of-bag data; only the winner advances. The simpler LGB^A+ variant alternates blocks of tree updates with a greedy linear correction. Because the final prediction is the sum of a linear component and a tree component, forecasts, permutation importance, and proximity weights decompose natively into linear and nonlinear contributions. In a quarterly U.S. macroeconomic forecasting exercise, LGB+ produces strong accuracy gains for targets that exhibit pronounced autoregressive dynamics or mixed linear-nonlinear signals, with variables entering the linear channel typically those that operate through persistence or near-accounted

What carries the argument

The LGB+ procedure that, at each boosting iteration, selects between a tree split and a linear coefficient update on the basis of out-of-bag performance (or alternates blocks of each), allowing the model to capture both linear autoregressive dynamics and nonlinearities without committing in advance to one functional form.

If this is right

Forecasts decompose additively into linear and nonlinear contributions without post-hoc approximation.
Permutation-based variable importance and historical proximity weights also separate cleanly into linear and tree channels.
Accuracy gains are largest for series with strong autoregressive persistence or mixed linear-nonlinear predictive content.
The procedure avoids ex-ante commitment to any particular functional form or predictor set.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The same selection logic could be applied to other time-series domains where linear trends coexist with nonlinear shocks.
Extending the candidate pool beyond linear and trees to additional simple basis functions might yield further efficiency gains in small samples.
Native linear-nonlinear decomposability could help isolate persistent versus transitory components in policy forecasts.

Load-bearing premise

Out-of-bag evaluation reliably chooses between linear and tree updates without introducing selection bias or overfitting to the particular macroeconomic series and time periods examined.

What would settle it

Re-estimating the same quarterly U.S. forecasting exercise on a later sample period or different set of target variables and finding no accuracy improvement over ordinary gradient boosting would falsify the performance claim.

read the original abstract

Needless to say, linear dynamics are pervasive in economic time series, particularly autoregressive ones. While gradient boosting with trees excels at capturing nonlinearities, it is inefficient in small samples when much of the predictive content is linear, expending splits to approximate relationships better captured by simple linear terms. This paper proposes LGB+, a boosting procedure operating on a more inclusive set of basis functions. The idea comes in two flavors. LGB+ evaluates a tree and a linear candidate at each step against out-of-bag data; only the winner advances. The simpler variant, LGB^A+, alternates on a fixed schedule: a block of tree updates, then a greedy linear correction, repeat. Both designs avoid ex ante commitments to any particular functional form or predictor selection. Because the prediction is the sum of a linear and a tree component, forecasts decompose natively into linear and nonlinear contributions, and so does permutation-based variable importance and historical proximity weights. In a quarterly U.S. macroeconomic forecasting exercise, LGB+ delivers strong gains for targets with pronounced autoregressive dynamics or mixed linear-nonlinear signals. Variables dominating the linear channel are those operating through autoregressive persistence or near-accounting relationships to the target (e.g., initial claims for unemployment and building permits for housing starts).

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

LGB+ sensibly mixes linear candidates into boosting for macro series with AR content, but OOB selection in dependent data likely inflates the claimed gains for exactly those targets.

read the letter

The main takeaway is a boosting design that lets linear terms compete with trees at each step or alternate in blocks, plus a native split of the forecast and importances into linear and nonlinear pieces. This directly addresses how trees waste splits approximating straight-line autoregressions that dominate many macro targets. The decomposition is a clean byproduct that helps see which variables operate through persistence versus other channels, as the abstract illustrates with initial claims and building permits. The approach avoids locking into one functional form upfront, which matches the mixed signals in quarterly U.S. data. That part of the work is straightforward and useful for applied forecasters. The soft spot sits in the validation. Selecting the winner by out-of-bag performance works when observations are independent, but quarterly macro series are serially correlated, so OOB draws share the same dependence structure as the training sample. This can make the selection favor whichever basis better fits the particular sample's autocorrelation rather than genuine predictive content, especially for the AR-heavy targets where gains are highlighted. The alternating variant does not escape the issue if it still relies on within-block OOB comparisons. The abstract states strong gains without numbers, baselines, or significance details, so the magnitude remains unclear until the full results and any time-series cross-validation checks are examined. This paper is aimed at macroeconomists and forecasters who already use boosting but want a lighter touch on linear structure. Readers running practical forecasting exercises would extract usable ideas from the design even if they adapt the selection rule. It deserves peer review because the algorithmic idea is clear and the application area matters, though the evaluation will need tightening to support the central claims.

Referee Report

3 major / 3 minor

Summary. The paper introduces LGB+ and LGB^A+, two variants of gradient boosting that augment standard tree updates with linear basis functions. LGB+ selects at each iteration between a tree candidate and a linear candidate by comparing their out-of-bag performance; LGB^A+ follows a fixed alternating schedule of tree blocks followed by a greedy linear correction. The resulting additive model permits native decomposition of forecasts, variable importance, and proximity weights into linear and nonlinear components. In a quarterly U.S. macroeconomic forecasting exercise the authors report that both variants deliver strong gains relative to standard tree boosting, especially for targets exhibiting pronounced autoregressive dynamics or mixed linear-nonlinear signals.

Significance. If the reported forecasting gains survive proper time-series validation, the procedure offers a practical, assumption-light way to blend linear and nonlinear modeling in macroeconometric forecasting while preserving interpretability through additive decomposition. The native linear/nonlinear split and the avoidance of ex-ante functional-form commitments are attractive features for applied work.

major comments (3)

[§3.2] §3.2 (OOB selection rule): The procedure selects between linear and tree updates by comparing out-of-bag performance. In quarterly macroeconomic series, observations are serially dependent, so standard OOB draws do not constitute independent validation sets. This risks the selection rule favoring whichever basis better fits the sample-specific serial-correlation pattern rather than genuine out-of-sample content, which directly undermines the central claim that gains are concentrated on AR-heavy targets.
[§4.2] §4.2 (forecasting exercise): The reported gains for targets with strong autoregressive dynamics rest on OOB-based model selection. No blocked or rolling-origin cross-validation results are presented to check whether the advantage survives when serial dependence is respected in the validation scheme. Without this check the quantitative superiority cannot be regarded as robust.
[Table 2 / Figure 3] Table 2 / Figure 3: The decomposition of variable importance into linear and nonlinear channels is presented, yet the paper does not report the exact linear specification (which predictors enter the linear term, whether lags are included, etc.). This makes it difficult to assess whether the linear channel is simply recovering standard AR terms or something more.

minor comments (3)

[§3.1] The notation distinguishing the linear component L_t from the tree component T_t could be introduced earlier and used consistently in the algorithmic description.
[Figure 1] Figure 1: Axis labels and legend entries for the linear versus nonlinear contributions are too small; the decomposition plots would be easier to read with larger fonts and explicit shading.
[Abstract] The abstract states that LGB+ 'delivers strong gains' but supplies no numerical magnitudes, baseline models, or significance levels; these details should appear in the abstract or be cross-referenced to the main tables.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the thoughtful and constructive report. The comments identify important issues regarding time-series dependence and interpretability that we address below. We propose targeted revisions to strengthen the robustness and transparency of the results while preserving the core contributions of LGB+ and LGB^A+.

read point-by-point responses

Referee: [§3.2] §3.2 (OOB selection rule): The procedure selects between linear and tree updates by comparing out-of-bag performance. In quarterly macroeconomic series, observations are serially dependent, so standard OOB draws do not constitute independent validation sets. This risks the selection rule favoring whichever basis better fits the sample-specific serial-correlation pattern rather than genuine out-of-sample content, which directly undermines the central claim that gains are concentrated on AR-heavy targets.

Authors: We agree that standard OOB sampling does not respect serial dependence and could in principle bias the basis-function selection toward in-sample autocorrelation patterns. At the same time, the OOB criterion is applied only internally at each boosting iteration to choose between a tree and a linear candidate; the final model is still evaluated on a genuine out-of-sample period with recursive estimation. To directly address the concern, the revised manuscript will report results under blocked and rolling-origin cross-validation for the selection step itself, allowing readers to see whether the reported advantage on AR-heavy targets survives when temporal ordering is strictly enforced. revision: partial
Referee: [§4.2] §4.2 (forecasting exercise): The reported gains for targets with strong autoregressive dynamics rest on OOB-based model selection. No blocked or rolling-origin cross-validation results are presented to check whether the advantage survives when serial dependence is respected in the validation scheme. Without this check the quantitative superiority cannot be regarded as robust.

Authors: The referee correctly notes that the main forecasting exercise relies on OOB for internal selection. We will add a new subsection (or appendix) that re-estimates the entire procedure under rolling-origin and blocked cross-validation schemes that preserve the time-series structure. These checks will be applied to the same set of targets, with particular attention to those exhibiting strong autoregressive dynamics. We expect the qualitative ranking to be preserved, but the quantitative magnitudes may be moderated; the revised tables will report both the original and the time-series-validated results side by side. revision: yes
Referee: [Table 2 / Figure 3] Table 2 / Figure 3: The decomposition of variable importance into linear and nonlinear channels is presented, yet the paper does not report the exact linear specification (which predictors enter the linear term, whether lags are included, etc.). This makes it difficult to assess whether the linear channel is simply recovering standard AR terms or something more.

Authors: We accept that greater transparency on the selected linear basis functions is needed. In the revision we will add an appendix table that, for each target, lists the linear terms retained by the greedy correction step (including the specific lags and predictors chosen). This will make explicit that the linear channel captures both autoregressive persistence and other near-accounting relationships (e.g., initial claims for unemployment, building permits for housing starts), while the nonlinear channel captures the residual interactions. The main text will reference this table when discussing Table 2 and Figure 3. revision: yes

Circularity Check

0 steps flagged

No significant circularity in LGB+ proposal or empirical claims

full rationale

The paper explicitly defines LGB+ as an algorithmic extension to gradient boosting that incorporates linear basis functions alongside trees, with selection performed either by out-of-bag comparison at each step or by fixed alternation in the simpler variant. The central claims concern observed forecasting gains on quarterly U.S. macroeconomic series for targets exhibiting autoregressive or mixed signals; these gains are obtained by applying the defined procedure to external data and reporting standard metrics. No derivation step reduces a result to its inputs by construction, no fitted parameters are relabeled as predictions, and no load-bearing self-citations or uniqueness theorems appear. The evaluation remains falsifiable on the held-out macro data, rendering the paper self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Based solely on the abstract, the paper introduces no explicit free parameters, axioms, or invented entities beyond the standard boosting framework; any hyperparameters (learning rate, tree depth, etc.) are left unspecified.

pith-pipeline@v0.9.0 · 5516 in / 1042 out tokens · 39073 ms · 2026-05-12T02:20:50.801525+00:00 · methodology

discussion (0)

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

IndisputableMonolith/Cost/FunctionalEquation.lean washburn_uniqueness_aczel unclear
LGB+ evaluates a tree and a linear candidate at each step against out-of-bag data; only the winner advances... the prediction is the sum of a linear and a tree component
IndisputableMonolith/Foundation/RealityFromDistinction.lean reality_from_one_distinction unclear
linear dynamics—autoregressive structures especially—are the dominant feature of most macroeconomic forecasting problems

Reference graph

Works this paper leans on

62 extracted references · 62 canonical work pages · 1 internal anchor

[1]

Instance-based Explanations for Gradient Boosting Machine Predictions with

Geertsema, Paul and Lu, Helen , journal=. Instance-based Explanations for Gradient Boosting Machine Predictions with

work page
[2]

Journal of the American Statistical Association , volume=

Forecasting Using Principal Components from a Large Number of Predictors , author=. Journal of the American Statistical Association , volume=

work page
[3]

Statistical models in S , pages=

Generalized additive models , author=. Statistical models in S , pages=. 2017 , publisher=

work page 2017
[4]

Journal of Business & Economic Statistics , volume=

Forecasting Inflation in a Data-Rich Environment: The Benefits of Machine Learning Methods , author=. Journal of Business & Economic Statistics , volume=

work page
[5]

Journal of Applied Econometrics , year=

The Macroeconomy as a. Journal of Applied Econometrics , year=

work page
[6]

Journal of the American Statistical Association , volume=

Local Linear Forests , author=. Journal of the American Statistical Association , volume=

work page
[7]

The Annals of Statistics , volume=

Generalized Random Forests , author=. The Annals of Statistics , volume=

work page
[8]

The Annals of Statistics , volume=

Greedy Function Approximation: A Gradient Boosting Machine , author=. The Annals of Statistics , volume=

work page
[9]

Computational Statistics & Data Analysis , volume=

Stochastic Gradient Boosting , author=. Computational Statistics & Data Analysis , volume=

work page
[10]

Boosting with the

B. Boosting with the. Journal of the American Statistical Association , volume=

work page
[11]

Chen, Tianqi and Guestrin, Carlos , booktitle=

work page
[12]

Ke, Guolin and Meng, Qi and Finley, Thomas and Wang, Taifeng and Chen, Wei and Ma, Weidong and Ye, Qiwei and Liu, Tie-Yan , booktitle=

work page
[13]

Journal of Applied Econometrics , volume=

How is Machine Learning Useful for Macroeconomic Forecasting? , author=. Journal of Applied Econometrics , volume=. 2022 , publisher=

work page 2022
[15]

Bayesian Modeling of Time-Varying Parameters Using Regression Trees , author=

work page
[16]

Blockwise Boosted Inflation: Nonlinear Determinants of Inflation Using Machine Learning , author=

work page
[17]

Nowcasting World Trade with Machine Learning: A Three-Step Approach , author=

work page
[18]

Statistical Science , volume=

Boosting Algorithms: Regularization, Prediction and Model Fitting , author=. Statistical Science , volume=

work page
[19]

Journal of Machine Learning Research , volume=

Model-based Boosting 2.0 , author=. Journal of Machine Learning Research , volume=

work page
[20]

The Annals of Statistics , volume=

Multivariate Adaptive Regression Splines , author=. The Annals of Statistics , volume=

work page
[21]

Quantifying the Risk-Return Tradeoff in Forecasting , author=

work page
[22]

Machine Learning , year=

To Bag is to Prune , author=. Machine Learning , year=

work page
[23]

and Ng, Serena , title =

McCracken, Michael W. and Ng, Serena , title =. 2020 , type =

work page 2020
[24]

and Watson, Mark W

Stock, James H. and Watson, Mark W. , title =. Journal of Business & Economic Statistics , year =

work page
[25]

Machine Learning , year =

Breiman, Leo , title =. Machine Learning , year =

work page
[27]

arXiv preprint arXiv:2601.06371 , year=

The Promise of Time-Series Foundation Models for Agricultural Forecasting: Evidence from Marketing Year Average Prices , author=. arXiv preprint arXiv:2601.06371 , year=

work page arXiv
[28]

Federal Reserve Bank of Philadelphia Research Rap Special Report , year =

Stark, Tom , title =. Federal Reserve Bank of Philadelphia Research Rap Special Report , year =

work page
[29]

and Williams, Jared , title =

Engelberg, Joseph and Manski, Charles F. and Williams, Jared , title =. Journal of Business & Economic Statistics , year =

work page
[30]

Journal of Business & Economic Statistics , volume=

A Neural. Journal of Business & Economic Statistics , volume=. 2025 , publisher=

work page 2025
[32]

International Journal of Forecasting , volume=

Macroeconomic Data Transformations Matter , author=. International Journal of Forecasting , volume=

work page
[33]

The Annals of Statistics , volume=

Least Angle Regression , author=. The Annals of Statistics , volume=. 2004 , publisher=

work page 2004
[34]

Athey, S., Tibshirani, J., and Wager, S. (2019). Generalized random forests. The Annals of Statistics , 47(2):1148--1178

work page 2019
[35]

Breiman, L. (2001). Random forests. Machine Learning , 45(1):5--32

work page 2001
[36]

Buckmann, M., Potjagailo, G., and Schnattinger, P. (2025). Blockwise boosted inflation: Nonlinear determinants of inflation using machine learning. Staff Working Paper 1143, Bank of England

work page 2025
[37]

and Hothorn, T

B \"u hlmann, P. and Hothorn, T. (2007). Boosting algorithms: Regularization, prediction and model fitting. Statistical Science , 22(4):477--505

work page 2007
[38]

and Yu, B

B \"u hlmann, P. and Yu, B. (2003). Boosting with the L2 loss: Regression and classification. Journal of the American Statistical Association , 98(462):324--339

work page 2003
[39]

and Guestrin, C

Chen, T. and Guestrin, C. (2016). XGBoost : A scalable tree boosting system. In Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining , pages 785--794

work page 2016
[40]

D., Meunier, B., and Stumpner, S

Chinn, M. D., Meunier, B., and Stumpner, S. (2023). Nowcasting world trade with machine learning: A three-step approach. Working paper, National Bureau of Economic Research

work page 2023
[41]

Efron, B., Hastie, T., Johnstone, I., and Tibshirani, R. (2004). Least angle regression. The Annals of Statistics , 32(2):407--499

work page 2004
[42]

F., and Williams, J

Engelberg, J., Manski, C. F., and Williams, J. (2009). Comparing the point predictions and subjective probability distributions of professional forecasters. Journal of Business & Economic Statistics , 27(1):30--41

work page 2009
[43]

Friedberg, R., Tibshirani, J., Athey, S., and Wager, S. (2021). Local linear forests. Journal of the American Statistical Association , 116(535):1448--1458

work page 2021
[44]

Friedman, J. H. (1991). Multivariate adaptive regression splines. The Annals of Statistics , 19(1):1--67

work page 1991
[45]

Friedman, J. H. (2001). Greedy function approximation: A gradient boosting machine. The Annals of Statistics , 29(5):1189--1232

work page 2001
[46]

Friedman, J. H. (2002). Stochastic gradient boosting. Computational Statistics & Data Analysis , 38(4):367--378

work page 2002
[47]

AXIL: Exact Instance Attribution for Gradient Boosting

Geertsema, P. and Lu, H. (2023). Instance-based explanations for gradient boosting machine predictions with AXIL weights. arXiv preprint arXiv:2301.01864

work page internal anchor Pith review Pith/arXiv arXiv 2023
[48]

Goulet Coulombe , P. (2024a). The macroeconomy as a R andom F orest. Journal of Applied Econometrics

work page
[49]

Goulet Coulombe , P. (2024b). To bag is to prune. Machine Learning . Forthcoming

work page
[50]

Goulet Coulombe , P. (2025). A neural P hillips curve and a deep output gap. Journal of Business & Economic Statistics , 43(3):669--683

work page 2025
[51]

Goulet Coulombe , P. (2026). Quantifying the risk-return tradeoff in forecasting. Working Paper

work page 2026
[52]

Goulet Coulombe , P., G \"o bel, M., and Klieber, K. (2024). The dual interpretation of machine learning forecasts. arXiv preprint arXiv:2412.13076

work page arXiv 2024
[53]

Goulet Coulombe , P., Leroux, M., Stevanovic, D., and Surprenant, S. (2021). Macroeconomic data transformations matter. International Journal of Forecasting , 37(4):1338--1354

work page 2021
[54]

Goulet Coulombe , P., Leroux, M., Stevanovic, D., and Surprenant, S. (2022). How is machine learning useful for macroeconomic forecasting? Journal of Applied Econometrics , 37(5):920--964

work page 2022
[55]

Hastie, T. J. (2017). Generalized additive models. Statistical models in S , pages 249--307

work page 2017
[56]

Hauzenberger, N., Huber, F., Koop, G., and Mitchell, J. (2023). Bayesian modeling of time-varying parameters using regression trees. Working paper, Federal Reserve Bank of Cleveland

work page 2023
[57]

Hollmann, N., M \"u ller, S., Eggensperger, K., and Hutter, F. (2022). TabPFN : A transformer that solves small tabular classification problems in a second. arXiv preprint arXiv:2207.01848

work page arXiv 2022
[58]

Hothorn, T., B \"u hlmann, P., Kneib, T., Schmid, M., and Hofner, B. (2010). Model-based boosting 2.0. Journal of Machine Learning Research , 11:2109--2113

work page 2010
[59]

Ke, G., Meng, Q., Finley, T., Wang, T., Chen, W., Ma, W., Ye, Q., and Liu, T.-Y. (2017). LightGBM : A highly efficient gradient boosting decision tree. In Advances in Neural Information Processing Systems , volume 30

work page 2017
[60]

and Medeiros, M

Masini, R. and Medeiros, M. (2025). Balancing flexibility and interpretability: A conditional linear model estimation via random forest. arXiv preprint arXiv:2502.13438

work page arXiv 2025
[61]

McCracken, M. W. and Ng, S. (2020). FRED-QD : A quarterly database for macroeconomic research. Working Paper 2020-005, Federal Reserve Bank of St. Louis

work page 2020
[62]

C., Vasconcelos, G

Medeiros, M. C., Vasconcelos, G. F. R., Veiga, A., and Zilberman, E. (2021). Forecasting inflation in a data-rich environment: The benefits of machine learning methods. Journal of Business & Economic Statistics , 39(1):98--119

work page 2021
[63]

Stark, T. (2010). Realistic evaluation of real-time forecasts in the S urvey of P rofessional F orecasters. Federal Reserve Bank of Philadelphia Research Rap Special Report , pages 1--20

work page 2010
[64]

Stock, J. H. and Watson, M. W. (2002a). Forecasting using principal components from a large number of predictors. Journal of the American Statistical Association , 97(460):1167--1179

work page
[65]

Stock, J. H. and Watson, M. W. (2002b). Macroeconomic forecasting using diffusion indexes. Journal of Business & Economic Statistics , 20(2):147--162

work page