The paper has a convincing definition of private-market peers and much thinner evidence that it selects them well.
Frank, Lyu, Jarmey, Saha, Li, Kaur, Akinola and Mehta begin with a global private-market universe of roughly 270,000 companies. Their sample keeps the roughly 53,000 firms with an observed or derivable post-money valuation and at least one PE or VC transaction. Each firm contributes one deal, its most recent. Deal history spans 1977 to 2025. The median deal age is five years, 75% of deals occurred since 2015, and the study treats the data as a cross-sectional snapshot from September 2025.
A CatBoost model predicts log post-money valuation from 27 features, including 23 categorical features. The target is winsorized at the 1st and 99th percentiles. An 80:20 split is stratified by internal subindustry, with recency weights applied to training samples. After 50 Optuna trials with five-fold CV, the selected model has depth 13 and 1,859 trees.
The fitted ensemble produces the similarity measure. Firms count as similar when they repeatedly reach the same terminal leaf, and each tree receives a weight based on its incremental reduction in training loss. Dissimilarity equals one minus weighted co-occurrence and is bounded in [0,1]. The authors call it a supervised dissimilarity because it is "not guaranteed to satisfy all metric-space axioms such as the triangle inequality."
Out-of-sample results show MAE of 1.08 against 1.19 for the OLS baseline, RMSE of 1.44 against 1.56, and R-squared of 0.46 against 0.36. The paper then turns to the evidence relevant to peer selection: 20-nearest-neighbor coherence and a weighted k-NN valuation horse race. Its baselines are cosine distance over MPNet embeddings of internal company descriptions, Euclidean distance over standardized and leave-one-out-encoded features, and Gower.
The paper links the method to valuation, diligence, portfolio construction and risk management. Better comparables could improve carrying values and identify holdings that appear diversified by industry while sharing pricing drivers. The paper contains no return, Sharpe or P&L, and makes no claim that it does.
What counts as a peer?
The central idea holds up. Fixed filters must trade coverage for relevance when the data contains 720 internal industry groups, 305 product and service categories, and 4,816 region-and-city values. Conditioning the partition on an observable outcome gives that trade-off an answer. In the authors' words, "companies are considered peers because they exhibit similar valuation dynamics given their underlying attributes." Identical valuations do not establish peer status when their drivers differ. That economic object is defensible, and terminal-leaf co-occurrence offers an inexpensive construction.
Deal taxonomy and geography run the map
SHAP shows a peer graph driven mainly by transaction taxonomy and geography. Deal type has mean absolute SHAP of 1.31, followed by country at 1.17. Region and city comes next at 0.28, ahead of revenue at 0.22 and average time since last financials at 0.20. The remaining 22 features contribute 1.62 combined. Missingness is heavy: 78% for revenue, 80% for EBITDA and 78% for net income.
Within Financial Services, the neighborhood test gives the learned metric lower mismatch rates on the top three categorical features. Those features are deal type, country, and region and city.
The result largely checks whether a metric built from those fields retrieves firms that match on those fields.
The comparison against the model's own numeric inputs says more. Again using Financial Services, the learned metric consistently beats cosine on the top three numeric features. Its record against Gower and Euclidean is mixed. Cosine similarity over company descriptions is the weakest baseline, while Gower and Euclidean remain competitive on numeric features.
This result also weakens the paper's most interesting proposed use: finding peers beyond conventional sector boundaries. A similarity map dominated by deal type and country, then evaluated within one industry group, has yet to demonstrate cross-sector analogues. The authors leave scalable cross-industry similarity for future work.
Does valuation consistency establish better peer selection?
The k-NN exercise is reasonable within its limits, and the paper addresses the immediate objection. It does not recycle the CatBoost point forecast. An independent non-parametric k-NN produces valuations from the neighborhood graph. Still, the metric learned from the target is asked to form neighborhoods that predict the same target, while cosine, Euclidean and Gower received no such supervision. Valuation consistency therefore gives a result close to confirming that training succeeded.
A practitioner needs a comparison with the method already in use. We did not find one against the CatBoost point prediction itself. We also did not find one against an industry-plus-stage-plus-geography median multiple.
The printed quantitative evidence is narrow beside the framework's stated scope. Figure 3 is captioned Financial Services (3.2k firms), and the text also places the Figure 4 k-NN comparison in Financial Services. Those two figures support the full baseline comparison. Only the direction appears in the text: the learned metric's MAE and RMSE are "consistently lower than the other metrics across all neighborhood sizes." It prints no MAE or RMSE gaps.
Both the abstract and conclusion limit the claim to "in the evaluated industry groups," an honest qualification. The conclusion adds another hedge, saying the metric "generally outperforms" the baselines. Yet the abstract presents that result alongside the scale of the full universe: 270,000 companies, more than 53,000 valuations, and coverage across multiple industries, geographies and deal stages. Evidence from a 3.2k-firm figure carries the claim attached to a 270,000-company universe.
Marks, tails and the available peer pool
Out-of-sample RMSE varies sharply across the distribution. It is 0.94 in Q8, rises to 2.35 in Q1 and reaches 2.05 in Q10, compared with 1.44 overall. Later-stage PE and "unspecified" deals have the highest errors; early-stage VC prices more cleanly. Private marks are most painful in those tails.
In log space, MAE of 1.08 corresponds to an average absolute error near a factor of three. Reported MAPE is 0.06 and MdAPE is 0.05. Reconciliation with a factor-of-three error seems possible only if percentage errors were calculated on log valuations instead of levels. I did not find a level-space translation of the errors in the text.
The retransformation correction expresses the same problem differently. Multiplicative factors estimated on training data are 1.74 for Series transactions, 2.19 for other transactions, and 5.06 for unspecified ones. The bucket with the worst errors also requires a factor of five. Bias remains unmodeled.
Conformal coverage behaves well: 93% at 95% nominal, 74% at 75%, and 60% at 60%. The 95% interval width flags outliers at 3.35 times the mean prediction. The conformal table reports a count of 53.8k, which corresponds to the full calibration universe instead of a 20% held-out slice, although the table does not identify its sample. Elsewhere, the sampling section states that test-set valuations "do not influence model fitting, neighbor selection, similarity weights, scaling-factor estimation, or conformal calibration."
The peer pool carries two further limits. Peers are selected from the roughly one firm in five that both transacted and disclosed. The test set is concentrated as well: its ten most represented countries account for about 90%, and its most prevalent transaction types account for about 83%.
I did not find a time-ordered split. The stated procedure uses an 80:20 random split stratified by subindustry, applied to a September 2025 snapshot with recency weights. Training and test deals consequently overlap in time. The paper leaves usefulness for pricing a round that has not happened yet untested.
We could not test any of this ourselves. Our data contains no private-company universe, funding-round or acquisition records, post-money valuations, deal types or deal stages. US-listed fundamentals cannot substitute for the label used to train this method.
Several industry groups with numerical k-NN tables would change my view. So would a train-on-deals-through-2021 and test-on-2022-to-2025 split. For now, the paper offers a well-argued definition of a peer, supported by printed evidence from one industry group.