{"id":"a7c5d700-f91b-4a34-bbd4-a76562319093","arxiv_id":"2608.12594","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Using leaf co-occurrence in a valuation-trained CatBoost model as a similarity metric gives lower k-NN valuation errors than Euclidean, Gower, or embedding distances in Financial Services, the only sector shown.","lead":"The paper trains a CatBoost model on about 53,000 private company valuations and defines a company similarity score from how often two companies end up in the same tree leaf. It reports that this valuation-aware similarity finds better comparable companies than ordinary distance or text-embedding measures in the one industry group shown.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Heterogeneous construction of the target valuation may let the learned similarity exploit how valuations were computed rather than economic peerhood; the improvement claim needs a direct-vs-derived robustness check.","rationale":"I agree with the Pith reader's weakest assumption. The manuscript's own Section 4.1.1 admits that the target is a mix of direct and formula-derived post-money valuations, and Section 3.1 claims valuations reflect investors' assessment only when the valuation is directly observed. The abstract's strongest claim inherits this risk directly. I considered instead whether the absence of a random-forest-proximity baseline makes the comparison unfair; that is a real but secondary issue, because even an unfair comparison would not invalidate the literal claim that the learned metric beats Euclidean, Gower, and cosine distances. Target-construction heterogeneity, by contrast, can make the learned similarity solve a proxy task: predicting how the valuation was computed. The proposed test is selected because it directly compares the model's performance when the anchor is unambiguous. No code or data is provided, which makes such a check impossible for external readers, but the authors can run it internally. The framework remains plausible and the evaluation is careful about not reusing CatBoost predictions, but the central comparison is not yet established beyond this confound. Therefore I would not change the conditional verdict: the concern is addressable and should be a condition for acceptance.","tokens_in":10075,"tokens_out":7400,"duration_ms":84220,"concrete_test":"Re-run the full CatBoost training and Section 6.2 k-NN evaluation on the subset of firms with directly observed post-money valuations only, using the same hyperparameters and the same Euclidean, Gower, and cosine baselines; if sample size permits, also run the derived-only subset separately. Additionally, fit the full model with an explicit valuation-source indicator and report its SHAP value. If the learned metric no longer beats the baselines on the direct-only set, or if the source indicator has high SHAP importance, the reported advantage is likely an artifact of heterogeneous target construction rather than valuation-anchored peerhood.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central k-NN claim (Section 6.2) rests on post-money valuation being a well-defined target. Section 4.1.1 constructs the target from at least four heterogeneous sources: direct post-money valuations, enterprise value, deal size divided by acquisition percentage, and pre-money plus deal size or offering amount. These formulas are not economically equivalent: enterprise value is not equity value, acquisition-percentage conventions vary across deal types, and deal size may or may not include secondary shares. Because the valuation source is correlated with deal type, geography, and stage, CatBoost can learn leaf structures that encode \"how the number was computed\" rather than what the company is worth. The learned leaf-co-occurrence similarity then groups firms by construction recipe, and the k-NN valuation experiment, which reuses the same target values as labels, can improve MAE/RMSE by matching construction noise rather than economic peerhood. The paper does not report the composition of direct versus derived targets, nor does it test whether the learned similarity is stable across valuation-source groups. This is a load-bearing threat to the central claim because it attacks the supervised anchor itself: if the target is not a consistent measure of value, then \"valuation-anchored similarity\" is anchored to a proxy that is partly about data provenance.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes a supervised similarity-learning framework for private-market peer identification. The authors train a CatBoost gradient-boosted tree model on log post-money valuations of roughly 53,000 private companies and define pairwise company similarity as the tree-importance-weighted leaf-node co-occurrence frequency across the ensemble (Eqs. 1-6). They evaluate the learned dissimilarity in k-NN valuation tasks against Euclidean, Gower, and MPNet-embedding cosine baselines, reporting consistently lower MAE/RMSE in Financial Services, and also report SHAP-based and neighborhood-consistency analyses. The central claim is that the valuation-anchored similarity improves downstream k-NN valuation accuracy in the evaluated industry groups while retaining interpretability.","tokens_in":10337,"tokens_out":2779,"duration_ms":24444,"significance":"If the central claim holds, the framework would be a practically useful and commercially relevant extension of tree-proximity methods to private-market equity, and the paper would provide a concrete, explainable alternative to coarse-rule peer selection. The paper builds on established techniques (CatBoost, leaf-node co-occurrence, conformal prediction, SHAP) rather than introducing new theory, but the application domain and the explicit construction of a reusable pairwise similarity matrix are of potential interest to q-fin audiences. Strengths: the framework is clearly specified in closed form; the paper is candid about some limitations (dissimilarity not a metric, evaluation restricted to one industry group shown); the target construction and data provenance are described; and the design separates model fitting from test-set use. The empirical case is not yet convincing because results are shown for a single visible industry group with no uncertainty quantification.","major_comments":[{"comment":"The central claim of improved downstream k-NN valuation over Euclidean, Gower, and MPNet baselines is supported only by a single visible figure for Financial Services (Figure 4). No other industry groups, no numeric MAE/RMSE values, no standard errors, and no significance tests are reported, so the stated claim that improvements hold 'in the evaluated industry groups' cannot be assessed. Please report per-industry-group numeric results with error bars or paired tests across all evaluated groups, or narrow the claim accordingly.","section":"Section 6.2 and Abstract"},{"comment":"The k-NN evaluation reuses as labels the same 53,000 log-valuation targets used to train the CatBoost model that defines the similarity metric (Eqs. 1-5). This is not a circular use of CatBoost predictions, but it does mean the comparison is supervised similarity against unsupervised baselines, which may largely reflect the value of supervision rather than the specific tree-weighting mechanism. To support the specific contribution, compare against an equally supervised baseline (e.g., random-forest proximity trained on the same target, or a metric learned by a deep metric-learning model), and report k-NN results with the tree-importance weights set to uniform to isolate the effect of Eq. 4.","section":"Sections 3.4.1-3.4.2 and 6.2"},{"comment":"The target variable is assembled from four heterogeneous sources - direct post-money valuation, enterprise value, deal size divided by acquisition percentage, and pre-money value plus deal size/offering amount - that are not economically equivalent. Because source type is correlated with deal type and stage, CatBoost can in principle pick up construction-recipe signatures rather than economic peerhood, and the learned similarity could then group companies by how their valuations were computed. Please report the share of targets from each source, test whether learned neighborhoods and k-NN gains are stable within source groups, and, ideally, retrain on direct post-money valuations only as a robustness check.","section":"Sections 2.1 and 4.1.1"}],"minor_comments":[{"comment":"The definition of training-loss improvement is a heuristic; note that because CatBoost uses ordered boosting with random permutations, training-loss differences are not variance-stabilized and the weights w_t can be dominated by noisy early iterations. A sentence acknowledging this and reporting the robustness of Eq. 5 to alternative weightings would strengthen the paper.","section":"Section 3.4.1, Eqs. 3-4"},{"comment":"The descriptions of multi-label concatenation and the cardinality for 'Region and City (4,816)' are unclear; specify how multi-label deal types are combined and how the region-and-city cardinality was counted.","section":"Section 4.1.2 and Table 2"},{"comment":"Recency-based sample weighting is described only qualitatively; provide the functional form of the weights or state explicitly that the weighting function is not disclosed.","section":"Section 4.2.2"},{"comment":"The retransformation scaling factors of 1.74, 5.06, and 2.19 by deal type are surprisingly large; briefly explain why 'Unspecified' transactions require a factor above 5 and whether these factors are stable across folds.","section":"Section 5.2"},{"comment":"The MAPE and MdAPE values are reported as percentages of log-space errors; clarify the definition because percentage errors are normally computed in level space, and the 0.05-0.07 values are not interpretable as conventional MAPE.","section":"Section 5.1 and Table 4"},{"comment":"The figure captions are too terse; Figure 3 and Figure 4 should state the plotted quantity, the number of companies, and the error metric used.","section":"Figures 3 and 4"},{"comment":"The claim that the learned metric shows 'mixed performance against Gower and Euclidean distances' on numerical features in Financial Services should be quantified or reconciled with the k-NN improvements in the same group.","section":"Section 6.1"},{"comment":"The paper cites arXiv preprint and SSRN references without DOIs or access dates; the reference list should be completed before publication.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The paper is from an industry group and reads like an application note. The main risk is not the framework definition, which is standard, but the evidentiary basis for the empirical claim: one visible industry group, no error bars, and a supervised-versus-unsupervised comparison that the text does not fully contextualize. The heterogeneous target-construction issue is worth raising as a robustness requirement rather than as grounds for rejection. If the authors supply the requested robustness and uncertainty analysis, the paper could become acceptable for the q-fin audience."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know two things about this paper. First, the core method is not new: importance-weighted leaf-node co-occurrence from trees is Breiman proximity with a training-loss weighting, and this same group has applied it to corporate and municipal bonds. The genuine new piece is the application to private-company valuations at scale, with a focus on producing a reusable similarity matrix rather than just a prediction model. Second, the central claim of improvement over unsupervised baselines is plausible but under-supported as presented. The evaluation shows one industry group (Financial Services, 3.2k firms), gives no error bars or significance tests, and leaves out key implementation details like how the k-NN weights are computed. That is a fixable problem, not a fatal one.\n\nThe paper does several things well. The framework is clearly formalized in Eqs. 1-6, and the choice of CatBoost is sensible for high-cardinality categoricals and missing data; the use of ordered target statistics and ordered boosting shows attention to leakage. The conformal prediction intervals and SHAP analysis are nice additions that give the work practical texture. The writing is honest about what the metric is and is not (it does not satisfy the triangle inequality), and the related work is properly placed.\n\nThe soft spots are real and need to be addressed before I would trust the headline conclusion. The circularity concern is the most serious: the similarity is learned from the same post-money valuations that the k-NN evaluation uses as labels. The authors argue they do not reuse CatBoost predictions, but that misses the point. The similarity structure itself is supervised by those labels, so beating generic unsupervised distances may just mean the tree model has memorized the target well. A proper test would compare against a supervised baseline trained on a shuffled or unrelated target, or use nested cross-validation where the similarity is re-computed in each fold. The target construction heterogeneity is also a load-bearing concern. Section 4.1.1 lists four different ways to derive the post-money valuation, and these are not economically equivalent: enterprise value is not equity value, acquisition-percentage conventions vary, and deal size can include secondary shares. Because the construction method correlates with deal type and stage, the model may learn to cluster firms by how their valuation was computed rather than by economic peerhood. The stress-test note is right to flag this. A robustness check splitting the sample by direct vs. derived targets would be straightforward and should be standard.\n\nThere are also minor issues: the tree importance weighting via absolute training-loss difference is a reasonable heuristic but not deeply motivated, and the paper shows only one industry group in the k-NN experiment despite claiming multiple. The citation pattern is fine; earlier tree-proximity work is properly acknowledged.\n\nThis paper is for practitioners in private-market analytics and researchers applying tree-based ML to finance. It deserves a serious referee, not a desk reject, because the problem is important and the method is a credible adaption of known machinery. But the referee should require the authors to supply the missing robustness checks, error estimates, and code or data synthesis before publication. I would send it out, expecting at least major revision.","headline":"A competent application of known tree-proximity methods to private-market peer identification, but the headline improvement claim is thinner than the abstract suggests and needs stronger empirical support.","tokens_in":10899,"tokens_out":2205,"would_cite":false,"duration_ms":26686,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Private-company peers are best defined by a valuation-trained tree ensemble: importance-weighted leaf co-occurrences form a similarity metric that improves k-NN valuation over Euclidean, Gower, and text-embedding baselines in the…","keywords":["private markets","company similarity","supervised similarity learning","gradient-boosted trees","leaf-node co-occurrence","k-nearest neighbor valuation","post-money valuation","peer identification"],"falsifier":"Train the identical ensemble and leaf-co-occurrence similarity on the subset of companies whose post-money valuations were directly observed, hold out the derived-valuation companies, and compare k-NN MAE and RMSE against the Euclidean, Gower, and text-embedding baselines; if the learned metric no longer dominates, or if nearest neighbors cluster by valuation-source type, the central claim is falsified.","tokens_in":9881,"feed_emoji":"📊","tokens_out":12009,"duration_ms":96391,"temperature":0.7,"pith_summary":"The paper argues that in private markets, where disclosure is sparse and transactions are infrequent, the meaning of \"peer company\" should be learned from market valuations rather than imposed by industry codes or text descriptions. It trains a gradient-boosted tree ensemble on 53,000 observed or derived post-money valuations and defines two companies as similar to the extent that the ensemble repeatedly routes them into the same terminal leaf, weighting each tree by its incremental reduction in training loss. That learned similarity is then used, independently of the model's own point predictions, to run k-nearest-neighbor valuation. In the evaluated industry groups the learned metric yields lower MAE and RMSE than Euclidean, Gower, and text-embedding distances, while keeping peer relationships explainable through the features that drive the splits. If the result holds, investors gain a data-driven, case-based way to define peer groups across sector, geographic, and stage boundaries.","feed_headline":"Learned similarity beats standard distances for private valuations","feed_subtitle":"Valuation-trained tree leaves define peers that beat standard distances in k-NN tests.","key_machinery":"The central object is the tree-importance-weighted leaf-node co-occurrence similarity computed from an ensemble of gradient-boosted trees trained on log post-money valuation. For two companies $X_1$ and $X_2$, with $Z_{1,t}$ and $Z_{2,t}$ their terminal leaves in tree $t$, the similarity is $S(X_1,X_2)=\\sum_t w_t\\,\\mathbb{I}[Z_{1,t}=Z_{2,t}]$, where $w_t$ is the normalized absolute reduction in training loss contributed by tree $t$; dissimilarity is $1-S(X_1,X_2)$. This turns the ensemble's splits into a pairwise geometry: firms are peers when valuation-relevant feature interactions repeatedly route them together, and the tree-boosting algorithm's handling of high-cardinality categorical variables and missing values makes the learned partitions feasible on sparse private-market data.","core_discovery":"The central claim is that company similarity in private markets is better modeled as an output of supervised valuation learning than as a static feature distance. After fitting a gradient-boosted tree model to log post-money valuation, the authors define the similarity of two companies as the importance-weighted frequency with which both are routed to the same terminal leaf across the ensemble; weights are each tree's absolute reduction in training loss. The complement of this score is a bounded pairwise dissimilarity that is symmetric but not necessarily metric. Using about 53,000 valuation targets drawn from a roughly 270,000-company global universe, the paper reports that k-nearest-neighbor valuation estimates built on the learned metric have lower MAE and RMSE than k-NN built on Euclidean, Gower, or text-embedding cosine distances in the evaluated industry groups, and that the neighborhoods are explainable through Shapley-value feature attributions.","pith_inferences":["An implication the authors leave implicit is that the same importance-weighted leaf co-occurrence construction transfers to any scarce-outcome private-market target, such as private-credit pricing or exit likelihood, whenever the target variable defines what economic peerhood should mean.","A testable diagnostic is to train the identical pipeline on only the directly observed post-money valuations and check whether the k-NN gains persist; if the metric no longer beats the baselines, the reported improvement is driven by construction patterns in derived valuation targets rather than economic content.","Because the training snapshot spans nearly fifty years with recency weighting, the learned geometry likely drifts with market conditions; a natural extension is to re-estimate the ensemble on rolling windows and compare k-NN error across vintages."],"forward_implications":["k-nearest-neighbor valuation built on the learned similarity achieves lower MAE and RMSE than k-NN built on Euclidean, Gower, or text-embedding distances in the evaluated industry groups, without reusing the ensemble's own point predictions.","The same learned similarity matrix can support sourcing, benchmarking, and concentration risk analysis: firms that share valuation drivers appear as neighbors even when conventional sector or geography filters would separate them.","Because the similarity derives from a fixed trained ensemble, new companies can be projected into the same leaf-co-occurrence geometry, so peer relationships remain explainable by the features that drive the splits.","Valuation uncertainty and peer structure can be reported coherently from one fitted model, since conformal prediction intervals and the similarity matrix are both outputs of the same ensemble."],"supporting_citations":[{"why":"introduces random-forest proximity, the conceptual origin of similarity as shared terminal nodes","marker":"[11]"},{"why":"supplies the tree-boosting algorithm with ordered categorical statistics and native missing-value handling used to build the ensemble","marker":"[19]"},{"why":"applies tree-based supervised similarity to corporate bonds, the direct methodological predecessor being extended to private equities","marker":"[13]"},{"why":"shows machine-learned peer weighting can improve relative valuation, anchoring the valuation-conditioned premise","marker":"[4]"},{"why":"applies tree-based relative valuation to illiquid municipal bonds, another precedent for valuation-driven proximity","marker":"[5]"},{"why":"provides the text-embedding company-similarity baseline the k-NN experiment must beat","marker":"[9]"},{"why":"formalizes tree-induced proximity as a supervised geometry, supporting the metric interpretation","marker":"[12]"},{"why":"supplies the distance-metric-learning result that task-specific similarity often beats generic geometric distance","marker":"[14]"}],"fun_headline_variants":["Valuation-trained peers beat standard distances in private KNN","Supervised similarity outperforms static distances for private valuations","Tree-leaf similarity wins over distance metrics for private firms","Private-market peers learned from valuations, not features"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes that the 53,000 valuations used for training are economically comparable even when some are directly reported and some are reconstructed from deal terms or enterprise values; if reconstruction noise is substantial, the learned peer groups may reflect how valuations were computed rather than why companies are economic peers.","fun_headline_variants_meta":{"raw":{"variants":["Valuation-trained peers beat standard distances in private KNN","Supervised similarity outperforms static distances for private valuations","Tree-leaf similarity wins over distance metrics for private firms","Private-market peers learned from valuations, not features"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000895,"raw_usage":{"total_tokens":3838,"prompt_tokens":906,"completion_tokens":2932,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":522,"completion_tokens_details":{"reasoning_tokens":2868}},"tokens_in":522,"tokens_out":2932,"duration_ms":20416,"temperature":1.0,"reasoning_tokens":2868,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T00:03:58.695654+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the identical ensemble and leaf-co-occurrence similarity on the subset of companies whose post-money valuations were directly observed, hold out the derived-valuation companies, and compare k-NN MAE and RMSE against the Euclidean, Gower, and text-embedding baselines; if the learned metric no longer dominates, or if nearest neighbors cluster by valuation-source type, the central claim is falsified.","supporting_citations":[{"cited_title":"Supervised similarity learning for corporate bonds using ran- dom forest proximities","cited_arxiv_id":null,"evidence_quote":"applies tree-based supervised similarity to corporate bonds, the direct methodological predecessor being extended to private equities"},{"cited_title":"Machine learning-based rela- tive valuation of municipal bonds","cited_arxiv_id":null,"evidence_quote":"applies tree-based relative valuation to illiquid municipal bonds, another precedent for valuation-driven proximity"},{"cited_title":"Company similarity using large language models","cited_arxiv_id":null,"evidence_quote":"provides the text-embedding company-similarity baseline the k-NN experiment must beat"},{"cited_title":"Geometry-and accuracy- preserving random forest proximities.IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(9):10947–10959, 2023","cited_arxiv_id":null,"evidence_quote":"formalizes tree-induced proximity as a supervised geometry, supporting the metric interpretation"},{"cited_title":"Distance metric learning with application to clustering with side-information","cited_arxiv_id":null,"evidence_quote":"supplies the distance-metric-learning result that task-specific similarity often beats generic geometric distance"}],"review_version":1}