10⁹ into 109. The quotes are reproduced exactly as the extraction left them and have NOT been repaired — the meaning of such a code point belongs to the one PDF it came from, and guessing it would produce a clean, confident, unverifiable quote.phenomenon · claim-type · cross-domain-transfer · mechanism · claim-relation · replication-status · limitation — grounded in the full text| phenomenon | trees and neural networks improve return prediction over linear and generalized linear models, traced to their allowance of interactions among predictors— “We find that trees and neural networks unambiguously improve return prediction with monthly stock-level R2's between 0.33% and 0.40%. But the generalized linear model, which introduces nonlinearity via spline functions of each individual baseline predictor (but with no predictor interactions), fails to robustly outperform the linear… | |
| claim-type | method— “We identify the best performing methods (trees and neural networks) and trace their predictive gains to allowance of nonlinear predictor interactions” | |
| method — evidence | — “We conduct a large scale empirical analysis, investigating nearly 30,000 individual stocks over 60 years from 1957 to 2016. Our predictor set includes 94 characteristics for each stock, interactions of each characteristic with eight aggregate time series variables, and 74 industry sector dummy variables, totaling more than 900 baseline… | |
| object of study — evidence | — “we conduct a comparative analysis of machine learning methods for finance. We do so in the context of perhaps the most widely studied problem in finance, that of measuring equity risk premia.” | |
| research question — evidence | — “We perform a comparative analysis of machine learning methods for the canonical problem of empirical asset pricing: measuring asset risk premia.” | |
| contribution — evidence | — “we provide a new set of benchmarks for the predictive accuracy of machine learning methods in measuring risk premia of the aggregate market and individual stocks. This accuracy is summarized two ways. The first is a high out-of-sample predictive R2 relative to preceding literature that is robust across a variety of machine learning… | |
| transfer | from neurobiology: the architecture and vocabulary of the biological nervous system - neurons, synapses, axons, signals passed between layers - carried into empirical asset pricing as the description of a feed-forward network— “Analogous to axons in a biological brain, layers of the networks represent groups of "neurons" with each layer connected by "synapses" that transmit signals among neurons of different layers.” | transfers[1] |
| mechanism | the sample is cut into three disjoint periods that keep the data's time ordering: the model is estimated on the training period alone, while the hyperparameters are chosen by scoring forecasts on the second period and re-estimating from the training data at every step— “In particular, we divide our sample into three disjoint time periods that maintain the temporal ordering of the data. The first, or "training," subsample is used to estimate the model subject to a specific set of tuning parameter values. The second, or "validation," sample is used for tuning the hyperparameters. We… | mechanisms[0] |
| mechanism | the third period is used for neither estimation nor tuning, which is what makes it the only one of the three that is truly out of sample - the validation fits are not, because they are an input to the estimation— “Thus the third, or "testing," subsample, which is used for neither estimation nor tuning, is truly out-of-sample and thus is used to evaluate a method's predictive performance.” | mechanisms[1] |
| mechanism | the out-of-sample R-squared is computed against a forecast of zero rather than against the historical mean return, because the historical mean of an individual stock is so noisy that benchmarking against it would lower the bar - and the paper states the size of the choice: every method's monthly R-squared rises by roughly three percentage points under the conventional denominator— “A subtle but important aspect of our R2 metric is that the denominator is the sum of squared excess returns without demeaning. In many out-of-sample forecasting applications, predictions are compared against historical mean returns. While this approach is sensible for the aggregate index or long-short portfolios, for example, it is flawed… | mechanisms[2] |
| mechanism | the Diebold-Mariano comparison is applied to the cross-sectional AVERAGE of each month's prediction errors rather than to individual stock errors, because the test's weak-dependence condition is unlikely to hold across a cross-section of stocks with potentially strong dependence— “While time series dependence in returns is sufficiently weak, it is unlikely that the conditions of weak error dependence underlying the Diebold-Mariano test apply to our stock-level analysis due of potentially strong dependence in the cross section. We adapt Diebold-Mariano to our setting by comparing the cross-sectional average of… | mechanisms[3] |
| mechanism | the network architectures are fixed in advance instead of searched over, because searching uncountably many is called unrealistic and unnecessary - and what the authors say they hope to obtain is a reasonable LOWER BOUND on machine-learning performance, not the best network— “It is unrealistic and unnecessary to find the optimal network by searching over uncountably many architectures. Instead, we fix a variety of network architectures ex ante and estimate each of these. What we hope to achieve is reasonably lower bound on the performance of machine learning methods.” | mechanisms[4] |
| mechanism | the squared-error objective is replaced by a hybrid of squared loss for relatively small errors and absolute loss for relatively large ones, with the crossover controlled by a tuning parameter optimised adaptively from the data— “The Huber loss, H(�), is a hybrid of squared loss for relatively small errors and absolute loss for relatively large errors, where the combination is controlled by a tuning parameter, , that can be optimized adaptively from the data.” | mechanisms[5] |
| mechanism | trees in the forest are de-correlated by offering each branch only a randomly drawn subset of the predictors, so that a single dominant predictor cannot force the same early split in every bootstrap sample and collapse the variance reduction that averaging is supposed to provide— “The forest method de-correlates trees using a method known as "dropout," which considers only a randomly drawn subset of predictors for splitting at each potential branch. Doing so ensures that, in the example, early branches for at least a few trees will split on characteristics other than firm size. This lowers the average… | mechanisms[6] |
| mechanism | the network is estimated from several different random initialisations and the forecasts are averaged, because the optimiser is stochastic and a single seed's forecast carries variance that has nothing to do with the data— “In particular, we use multiple random seeds to initialize neural network estimation and construct predictions by averaging forecasts from all networks. This reduces prediction variance because the stochastic nature of the optimization can cause different seeds to produce different forecasts.” | mechanisms[7] |
| mechanism | the interpretation of which methods win is checked against two simulated data sets whose generating processes are known - one linear and additive, one nonlinear with pairwise interactions - so that the ranking can be shown to track the property it is being credited to— “We apply machine learning to simulated data from two different data generating processes. Both produce data from a high dimensional predictor set. But in one, individual predictors enter only linearly and additively, while in the other predictors can enter through nonlinear transformations and via pairwise interactions. When we apply our… | mechanisms[8] |
| qualifies | Green et al. (2013), Harvey et al. (2016), Welch and Goyal (2008) — the predictor-counting literature— “Green et al. (2013) count 330 stock-level predictive signals in published or circulated drafts. Harvey et al. (2016) study 316 "factors," which include firm characteristics and common factors, for describing stock return behavior. They note that this is only a subset of those studied in the literature. Welch and Goyal (2008) analyze… | claim_relations[0] |
| builds-on | Welch and Goyal (2008)— “As a benchmark for comparison, nearly all of the macroeconomic return predictor variables in the survey of Welch and Goyal (2008) fail to produce a positive out-of-sample forecast R2.” | claim_relations[1] |
| distinguishes-from | principal components regression (PCA)— “PCA condenses data into components based on the covariation among the predictors. This happens prior to the forecasting step and without consideration of how predictors associate with future returns. In contrast, partial least squares performs dimension reduction by directly exploiting covariation of predictors with the forecast target.” | claim_relations[2] |
| sense | out-of-sample R2 - denominator is the sum of squared excess returns, NOT demeaned, so the benchmark forecast is zero— “A subtle but important aspect of our R2 metric is that the denominator is the sum of squared excess returns without demeaning. In many out-of-sample forecasting applications, predictions are compared against historical mean returns. While this approach is sensible for the aggregate index or long-short portfolios, for example, it is flawed… | senses[0] |
| sense | out-of-sample R2 - the conventional sense, benchmarked against historical mean returns, which the paper says would raise all its numbers by about three percentage points— “In many out-of-sample forecasting applications, predictions are compared against historical mean returns. While this approach is sensible for the aggregate index or long-short portfolios, for example, it is flawed when it comes to analyzing individual stock returns. Predicting future excess stock returns with historical averages typically… | senses[1] |
| sense | machine learning - the authors' stipulated three-part sense: high-dimensional prediction models, regularization, and search algorithms— “We use the term to describe (i) a diverse collection of high-dimensional models for statistical prediction, combined with (ii) so-called "regularization" methods for model selection and mitigation of overfit, and (iii) efficient algorithms for searching among a vast number of potential model specifications.” | senses[2] |
| sense | risk premium - used interchangeably with expected return, while the paper notes components one may wish to distinguish— “We use the terms "expected return" and "risk premium" interchangeably. One may be interested in potentially distinguishing among different components of expected returns such as those due to systematic risk compensation, idiosyncratic risk compensation, or even due to mispricing.” | senses[3] |
| sense | variable importance - a ranking used to identify influential covariates, explicitly not an account of mechanism— “Our goal in interpreting machine learning models is modest. We aim to identify covariates that have an important influence on the cross-section of expected returns while simultaneously controlling for the many other predictors in the system.” | senses[4] |
| control | parameter penalisation and deliberately conservative model-selection criteria applied across the whole method comparison — held — rules out overfitting and false discovery as the source of apparent predictive power - the flexible functional forms being compared can fit noise, and without the penalty the comparison would reward whichever method fits it best— “parameter penalization and conservative model selection criteria complement the breadth of functional forms spanned by these methods in order to avoid overfit biases and false discovery” | controls[0] |
| differs in | Green et al. (2013), Harvey et al. (2016), Welch and Goyal (2008) — the predictor-counting literature — WHETHER THE COUNT IS OF PREDICTORS OR OF REPORTED PREDICTORS. Three surveys are cited with their totals - 330 stock-level signals, 316 factors, nearly 20 aggregate-market predictors - and then: 'there presumably exists a much larger set of predictors that were tested but failed to predict returns and were thus never reported.' The denominator of the search is unobservable, so the counts bound the published literature and not the work done— “Green et al. (2013) count 330 stock-level predictive signals in published or circulated drafts. Harvey et al. (2016) study 316 "factors," which include firm characteristics and common factors, for describing stock return behavior. They note that this is only a subset of those studied in the literature. Welch and Goyal (2008) analyze…— “there presumably exists a much larger set of predictors that were tested but failed to predict returns and were thus never reported” | claim_relations[0] |
| differs in | Welch and Goyal (2008) — WHAT THE BENCHMARK IS. The comparison point is not a rival method's success but a documented FAILURE: 'nearly all of the macroeconomic return predictor variables in the survey of Welch and Goyal (2008) FAIL TO PRODUCE A POSITIVE OUT-OF-SAMPLE FORECAST R2.' A negative result is doing the work of a baseline, which is only legible if the reader knows the sign convention of this R2 - and the record already carries two senses of it— “As a benchmark for comparison, nearly all of the macroeconomic return predictor variables in the survey of Welch and Goyal (2008) fail to produce a positive out-of-sample forecast R2.”— “fail to produce a positive out-of-sample forecast R2” | claim_relations[1] |
| differs in | principal components regression (PCA) — WHETHER THE DIMENSION REDUCTION KNOWS WHAT IT IS PREDICTING. 'PCA condenses data into components based on the covariation among the predictors. This happens prior to the forecasting step and without consideration of how predictors associate with future returns. In contrast, partial least squares performs dimension reduction by directly exploiting covariation of predictors with the forecast target.' Two reductions of the same data differing only in whether the target is in scope when the axes are chosen— “PCA condenses data into components based on the covariation among the predictors. This happens prior to the forecasting step and without consideration of how predictors associate with future returns. In contrast, partial least squares performs dimension reduction by directly exploiting covariation of predictors with the forecast target.”— “This happens prior to the forecasting step and without consideration of how predictors associate with future returns” | claim_relations[2] |
| edge | risk premium —same-as→ expected return— “We use the terms "expected return" and "risk premium" interchangeably.” | concept_edges[0] |
| evidence status note | predictive performance —operationalised-as→ out-of-sample R2 — This quote ends on 'as' because what it introduces is a displayed formula that the staged extraction does not carry as readable text (the text continues: "Roos 2 = 1 - (i,t)T3 (ri,t+1 - ri,t+1)2 (i,t)T3 "). The quote is verbatim and the formula is in the paper; the quote was not extended, because extending it would put extraction debris inside a quotation.— “To assess predictive performance for individual excess stock return forecasts, we calculate the outof-sample R2 as” | concept_edges[1] |
| edge | machine learning —contrasted-with→ traditional econometric prediction techniques— “The high-dimensional nature of machine learning methods (element (i) of this definition) enhances their flexibility relative to more traditional econometric prediction techniques” | concept_edges[4] |
| evidence status note | concept — This quote ends on 'as' because what it introduces is a displayed formula that the staged extraction does not carry as readable text (the text continues: "Roos 2 = 1 - (i,t)T3 (ri,t+1 - ri,t+1)2 (i,t)T3 "). The quote is verbatim and the formula is in the paper; the quote was not extended, because extending it would put extraction debris inside a quotation.— “To assess predictive performance for individual excess stock return forecasts, we calculate the outof-sample R2 as” | concepts[11] |
| outcome (stated by the tagger, no quotation) | control — held | controls[0] |
| for which use | limitation — comparing this paper's R2 against an R2 computed the usual way, against historical mean returns: the paper says that convention raises all its numbers by roughly three percentage points— “A subtle but important aspect of our R2 metric is that the denominator is the sum of squared excess returns without demeaning. In many out-of-sample forecasting applications, predictions are compared against historical mean returns. While this approach is sensible for the aggregate index or long-short portfolios, for example, it is flawed… | limitations[0] |
| for which use | limitation — reading the results as evidence about economic mechanisms: the paper says the measurements do not tell us about mechanisms or equilibria— “But these improved predictions are only measurements. The measurements do not tell us about economic mechanisms or equilibria. Machine learning methods on their own do not identify deep fundamental associations among asset prices and conditioning variables.” | limitations[1] |
| for which use | limitation — reading the comparison as covering the machine learning toolkit: the paper says it is not exhaustive and names an exclusion— “This is not an exhaustive analysis of all methods. For example, we exclude support vector machines as these share an equivalence with other methods that we study4 and are primarily used for classification problems.” | limitations[2] |
| for which use | limitation — reading the variable-importance results as an account of why a predictor works: the stated goal is only to identify influential covariates— “Our goal in interpreting machine learning models is modest. We aim to identify covariates that have an important influence on the cross-section of expected returns while simultaneously controlling for the many other predictors in the system.” | limitations[3] |
| for which use | limitation — the reported model-comparison p-values as standard Diebold-Mariano results: the test was adapted because its conditions were judged unlikely to hold— “it is unlikely that the conditions of weak error dependence underlying the Diebold-Mariano test apply to our stock-level analysis due of potentially strong dependence in the cross section. We adapt Diebold-Mariano to our setting by comparing the cross-sectional average of prediction errors from each model, instead of comparing errors among… | limitations[4] |
| for which use | limitation — the 900+ signal set as the space of tested predictors: the paper says an unknown larger set failed and was never reported— “In both stock and aggregate return predictions, there presumably exists a much larger set of predictors that were tested but failed to predict returns and were thus never reported.” | limitations[5] |
| for which use | limitation — generalising shallow-beats-deep beyond this setting: the paper attributes it, hedged, to the small data and low signal-to-noise ratio here— “This is likely an artifact of the relatively small amount of data and tiny signal-to-noise ratio for our return prediction problem, in comparison to the kinds of non-financial settings in which deep learning thrives thanks to astronomical datasets and strong signals (such as computer vision).” | limitations[6] |
| for which use | limitation — treating the target as an observable quantity: the conditioning information set is stated to be unobservable— “It amounts to best approximating the conditional expectation E(ri,t+1|Ft), where ri,t+1 is an asset's return in excess of the risk-free rate, and Ft is the true and unobservable information set of market participants.” | limitations[7] |
| for which use | limitation — reading "machine learning" here as a standard term: the paper says the definition is inchoate and stipulates its own three-part one— “The definition of "machine learning" is inchoate and is often context specific. We use the term to describe (i) a diverse collection of high-dimensional models for statistical prediction, combined with (ii) so-called "regularization" methods for model selection and mitigation of overfit, and (iii) efficient algorithms for… | limitations[8] |
| evidence status note | mechanism — 2026-08-16: THE EXTRACTION HAS STRIPPED THE MATH SYMBOLS INSIDE THIS QUOTE. It reads "The Huber loss, H(?), is a hybrid ... controlled by a tuning parameter, , that can be optimized adaptively" - the argument of H and the NAME OF THE TUNING PARAMETER are both absent from the staged file, leaving a bare comma where the symbol was. The quote is an exact substring and the file is incomplete here. The tag deliberately describes the parameter by its role rather than naming it, because naming it would mean supplying a character the staged text does not contain.— “The Huber loss, H(�), is a hybrid of squared loss for relatively small errors and absolute loss for relatively large errors, where the combination is controlled by a tuning parameter, , that can be optimized adaptively from the data.” | mechanisms[5] |
| limitation | The headline R2 uses a denominator without demeaning; the paper states the common alternative convention raises every method's R2 by roughly three percentage points.— “A subtle but important aspect of our R2 metric is that the denominator is the sum of squared excess returns without demeaning. In many out-of-sample forecasting applications, predictions are compared against historical mean returns. While this approach is sensible for the aggregate index or long-short portfolios, for example, it is flawed… | limitations[0] |
| limitation | The paper states its improved predictions are measurements only, and do not identify economic mechanisms or equilibria.— “But these improved predictions are only measurements. The measurements do not tell us about economic mechanisms or equilibria. Machine learning methods on their own do not identify deep fundamental associations among asset prices and conditioning variables.” | limitations[1] |
| limitation | The set of methods compared is explicitly not exhaustive, and support vector machines are excluded with a stated reason.— “This is not an exhaustive analysis of all methods. For example, we exclude support vector machines as these share an equivalence with other methods that we study4 and are primarily used for classification problems.” | limitations[2] |
| limitation | The paper describes its own interpretive goal as modest.— “Our goal in interpreting machine learning models is modest. We aim to identify covariates that have an important influence on the cross-section of expected returns while simultaneously controlling for the many other predictors in the system.” | limitations[3] |
| limitation | The conditions underlying the Diebold-Mariano test are stated to be unlikely to hold for the stock-level analysis; the paper modifies the test rather than proceeding.— “it is unlikely that the conditions of weak error dependence underlying the Diebold-Mariano test apply to our stock-level analysis due of potentially strong dependence in the cross section. We adapt Diebold-Mariano to our setting by comparing the cross-sectional average of prediction errors from each… | limitations[4] |
| limitation | The predictor literature the study draws on is stated to exclude an unknown larger set of predictors that were tested, failed and never reported. | limitations[5] |
| limitation | The finding that shallow networks beat deeper ones is attributed, with hedging, to the data available rather than established as a property of the problem. | limitations[6] |
| limitation | The estimation target is defined in terms of an information set the paper calls true and unobservable. | limitations[7] |
| limitation | The paper's central term is stated to have no settled definition, and the definition used is stipulated by the authors. Quotation withheld — this paper is not ours to reproduce, and the page has already quoted as much of it as the rights layer allows. The tag, its field and its confidence stand; the sentence is at the publisher. | limitations[8] |
| concept | machine learning Quotation withheld — this paper is not ours to reproduce, and the page has already quoted as much of it as the rights layer allows. The tag, its field and its confidence stand; the sentence is at the publisher. | concepts[1] |
| concept | risk premium Quotation withheld — this paper is not ours to reproduce, and the page has already quoted as much of it as the rights layer allows. The tag, its field and its confidence stand; the sentence is at the publisher. | concepts[2] |
| concept | expected return Quotation withheld — this paper is not ours to reproduce, and the page has already quoted as much of it as the rights layer allows. The tag, its field and its confidence stand; the sentence is at the publisher. | concepts[4] |
| concept | neural networks Quotation withheld — this paper is not ours to reproduce, and the page has already quoted as much of it as the rights layer allows. The tag, its field and its confidence stand; the sentence is at the publisher. | concepts[5] |
| concept | regression trees Quotation withheld — this paper is not ours to reproduce, and the page has already quoted as much of it as the rights layer allows. The tag, its field and its confidence stand; the sentence is at the publisher. | concepts[6] |
| concept | elastic net Quotation withheld — this paper is not ours to reproduce, and the page has already quoted as much of it as the rights layer allows. The tag, its field and its confidence stand; the sentence is at the publisher. | concepts[7] |
| concept | support vector machines Quotation withheld — this paper is not ours to reproduce, and the page has already quoted as much of it as the rights layer allows. The tag, its field and its confidence stand; the sentence is at the publisher. | concepts[8] |
| concept | variable importance Quotation withheld — this paper is not ours to reproduce, and the page has already quoted as much of it as the rights layer allows. The tag, its field and its confidence stand; the sentence is at the publisher. | concepts[9] |
| concept | out-of-sample R2 Quotation withheld — this paper is not ours to reproduce, and the page has already quoted as much of it as the rights layer allows. The tag, its field and its confidence stand; the sentence is at the publisher. | concepts[10] |
| concept | predictive performance Quotation withheld — this paper is not ours to reproduce, and the page has already quoted as much of it as the rights layer allows. The tag, its field and its confidence stand; the sentence is at the publisher. | concepts[11] |
| concept | traditional econometric prediction techniques Quotation withheld — this paper is not ours to reproduce, and the page has already quoted as much of it as the rights layer allows. The tag, its field and its confidence stand; the sentence is at the publisher. | concepts[12] |
| concept | Diebold-Mariano test Quotation withheld — this paper is not ours to reproduce, and the page has already quoted as much of it as the rights layer allows. The tag, its field and its confidence stand; the sentence is at the publisher. | concepts[13] |
When we have changed our mind. Some entries in the table begin with a word in capitals and the reason for it. EVIDENCE CORRECTED — the finding still stands, but we had quoted the wrong sentence from the paper, and the quotation has been replaced with the right one. RETRACTED — the entry itself was wrong; it stays on the page, marked, with the reason, rather than being deleted. WITHDRAWN AND REPLACED — the entry was wrong, and a corrected version of it appears in the entries below. Each mark carries its own reason, in brackets, because the reasons differ from one entry to the next. An entry we have WITHDRAWN is removed from this page and kept in the record, with its reason. A line reading Audit and a date is a later re-reading of the entry against its whole sentence: DOUBTFUL means it was kept, and the reason says what is loose about it. Quotation marks inside the tagger's sentence. Where a tag is the tagger's own sentence, a phrase in quotation marks inside it is borrowed from the paper — the tagger's wording around it is not. The paper's own sentence is the evidence quote for that row; on this page it is withheld with the article, and the article is at its publisher. Nothing inside quotation marks anywhere on this page is ours: if it is quoted, it is the paper's. Whose words a concept carries. A concept entry records a term the paper itself uses, so the term, the sense given for it, and its evidence are the paper's own words, quoted where the paper fixes them. This is the opposite of a sense entry in the tag table, which is the tagger's sentence about how the paper uses a term. Where a concept entry could not be grounded in the paper's words, the entry says so on its own row.
Tag layer © the Meta-Tagging Project, CC BY 4.0. The paper text is © its authors/publisher
and is linked, not reproduced. Excerpts shown are short quotations for identification and verification.
· back to index
· concepts · authors
· עברית
© 2026 Shir Sivroni · ORCID 0009-0000-2597-2824