PaperlayerView exact revision on arXiv

Source-linked AI summary

From language-model stock rankings to testable economic rules: A computational audit

Shuai Wu, Xue Li, Zhijun Wang, Bolun Liu, Weilin Cai, Zihao Su, Ran Wang

arXiv:2610.01213v1cs.CE

TL;DR

Language-model stock rankings can vary under identical inputs, and returns alone do not reveal the choices behind them. This audit tests ranking stability and frozen-rule reproduction; the rules approximate aggregate ranking responses, while primary return tests yield no adjusted finding.

  • Problem

    Returns alone reveal little about the choices behind language-model rankings, which may vary under identical inputs or respond unpredictably to input changes.

    Source: §1 · p. 1

  • Method

    Repeated requests measure decision variation, frozen rules test reproducible numerical structure, and matched portfolios measure returns associated with preferences beyond those rules.

    Source: §5 · p. 17

  • Results

    Frozen rules reach 0.923-0.984 SSE 50 and 0.795-0.985 transferred rank correlations, improve aggregate rank-change prediction, and show no adjusted finding in primary return tests.

    Source: §1 · p. 2, §5 · p. 17

  • Takeaways & Limitations

    The tests support approximation of aggregate ranking responses more strongly than exact selection changes.

    Source: §5 · p. 17

    The factor-controlled partial association is a conditional diagnostic, not a measure of coefficient-vector quality or a newly fitted trading strategy.

    Source: §3.4 · p. 6

Abstract

from arXiv · show

We test the stability, reproducibility and investment outcomes of language-model stock rankings. Four models and five numerical comparators share a portfolio engine over 72 monthly holding periods in the Shanghai Stock Exchange (SSE) 50, China Securities Index (CSI) 300 and CSI 500. Rankings use nine characteristics, and five repeated SSE 50 runs measure variation under identical inputs. Linear rules fitted to development-period model preferences are frozen before unseen-month, larger-pool and controlled-intervention tests. Their mean Spearman agreement with model rankings is 0.923-0.984 in the SSE 50 and 0.795-0.985 after transfer. Aggregate rank-change error falls relative to a zero-change prediction in twelve archived feature-group comparisons and eight matched single-feature comparisons, with Holm adjustments applied in separate nine-plus-three and six-plus-two families. Prediction of individual entries and exits remains weak (event Jaccard 0.000-0.125). Historical mean model compound annual growth rates range from 6.26% to 11.17%. At 10 basis points per side and six-month blocks, the twelve-comparison model-minus-rule return family and factor-controlled associations yield no adjusted finding. Three higher-cost, twelve-month-block comparisons favor a Terra rule within their twelve-test slices, with no adjusted finding across the full 144-test sensitivity grid. Two input-intervention batches totaling 5,184 responses supply paired intervention-return tests. The three-model batch has no bootstrap-adjusted finding at the primary block length; a Luna row-order effect appears under heteroskedasticity- and autocorrelation-consistent (HAC) adjustment within its three-test family but not in a pooled 24-test adjustment. Compact rules approximate aggregate rankings; return conclusions depend on comparison families, uncertainty methods and tie priorities.

1 Introduction

The paper audits the stability, reproducibility, responsiveness to controlled input changes, and economic value of language-model stock rankings, using frozen rules and repeated input interventions to examine aggregate rankings, rank changes, and exploratory return effects. It also describes archived materials for numerical reproduction and AI assistance with author review.

2.1 Data, samples and timing

  • The audit uses four models, three Chinese equity universes, nine shared contemporaneous characteristics, and 72 test months, with frozen development-period rules and controlled feature interventions.Development signals span December 2014-November 2019 and test signals December 2019-November 2025; the historical evaluation is ex post, while intervention inputs and final inference specifications were fixed before responses were collected.Source: §1 · p. 2, §2.1 · pp. 2–3, §2.2 · p. 3

3.2 Frozen-rule reproduction and transfer

  • Frozen linear rules reached SSE 50 rank correlations of 0.923-0.984 and correlations of 0.795-0.985 in larger stock pools.The frozen rules were transferred to the CSI 300 and CSI 500 without refitting; Spearman correlation is not a fraction of explained decisions or variance.Source: §1 · p. 2, §3.2 · p. 6

3.3 Controlled response tests

  • Aggregate rank-change prediction error fell in all twelve feature-group and eight single-feature comparisons, while individual entry/exit prediction remained weak (event Jaccard 0.000-0.125).Holm adjustments were applied in separate nine-plus-three and six-plus-two families; the matched single-feature experiment included 864 responses.Source: §1 · p. 2

3.1 Decision stability and return benchmark

  • At 10 basis points per side and six-month blocks, the twelve-comparison model-minus-rule return family and factor-controlled associations had no adjusted finding.The benchmark compares paired monthly net returns against equal-factor selection; three Terra effects at higher costs and twelve-month blocks were exploratory sensitivity results.Source: §1 · p. 2, §3.1 · p. 5

2.3 Numerical comparators and model requests

  • Numerical comparators include equal-factor and factor-category selection, while supervised ridge and gradient-boosting models predict next-period return ranks using expanding training samples.The factor-category comparator equally weights six categories; model requests use anonymous identifiers and provide characteristic definitions and a shared numerical matrix.Source: §2.3 · pp. 3–4

2.4 Portfolio construction and execution

  • Simulated returns incorporate costs, unfilled orders, cash, and retained corporate-event claims, while fractional positions and fixed notional costs simplify execution.The principal cost is 10 basis points per side, and round-lot sizing and time-varying market impact are outside the simulation.Source: §2.4 · p. 4

3.4 Economic value beyond the frozen rule

  • Beyond return comparisons, the study tests whether model preferences beyond the frozen rule correlate with forward returns, treating partial correlations as conditional diagnostics rather than a new strategy.The analysis projects preferences and returns off an intercept and nine factor columns; Holm adjustment is applied within declared test families.Source: §3.4 · pp. 6–7

3.5 Robustness and precision

  • Robustness exercises use risk-matched portfolios and disjoint-stock diagnostics, while separate precision calculations govern return tests and matched response experiments.The return calculation selected 36 months after the 24-month option missed its planned power target; the matched-response calculation selected 24 stratified months.Source: §3.5 · p. 7

3.6 Repeated input-intervention design

  • Astra contributed 1,296 responses and Luna, Terra, and Sol contributed 3,888 across six intervention conditions covering 72 SSE 50 test months.Each model had three complete runs per condition; the two batches shared the condition design and input matrices.Source: §3.6 · p. 7
  • Intervention returns were paired with replication primary-condition returns and tested by resampling paired trajectories and circular month blocks, with six-month blocks primary.The analyses used 10,000 draws, three-lag HAC tests, separate Holm-adjusted families, and a primary cost of 10 basis points per side; frozen-rule tests used other experiments.Source: §3.6 · p. 7

3.7 Reproducibility and artificial intelligence assistance

  • The research archive retains inputs, prompts, responses, frozen coefficients, portfolio records, and specifications for numerical reproduction without new model requests.AI tools assisted design, implementation, analysis, and writing; Shuai Wu led integration and checked the analyses, citations, and final text, and all authors reviewed their areas of expertise.Source: §3.7 · p. 7

4 Results

Frozen linear rules closely reproduce model rankings and predict aggregate responses to numerical changes, while individual selection changes remain difficult to predict. At the primary trading cost and block length, adjusted tests do not establish return advantages over equal-factor or frozen-rule portfolios.

4.1 Stability and reproducibility of ranking policies

  • Repeat overlap is 83.36%–94.63%; frozen-linear rank correlation is 0.923–0.984 in SSE 50 and 0.795–0.985 after transfer.Identical-input rankings are highly repeatable, though not deterministic; linear rules have higher rank agreement with choices than the bounded forest.Source: Table 3 · p. 8, §4.1 · pp. 8–9, Table 4 · p. 8

4.2 Prediction of controlled responses

  • All twelve archived numerical contrasts show positive MSE improvement and pass the adjusted 0.05 criterion under six-month blocks.The declared Holm families are separate: nine historical comparisons and three Astra comparisons.Source: Table 6 · p. 10, §4.2 · p. 10
  • Matched single-feature MSE improvements are 0.174 to 1.245, with all eight comparisons passing adjusted p below 0.05.Frozen predictions reduce aggregate squared rank-change error by 5.6%–55.7% versus zero-change, but event Jaccard is only 0.000–0.125 for entries and exits.Source: Table 7 · p. 10, §4.2 · p. 10

4.3 Investment performance under common execution

  • Mean model CAGR ranges 6.26%–11.17%, but none of twelve model-minus-equal-factor tests passes Holm adjustment at 10 basis points per side.The intervals are too wide to establish equivalence within the specified margin.Source: Table 8 · p. 11, §4.3 · pp. 11, 14, Table 9 · p. 13

4.4 Economic value beyond the frozen rule

  • At 10 basis points per side and six-month blocks, none of twelve model-minus-rule return tests meets the Holm-adjusted 0.05 criterion.Three Terra comparisons meet the criterion within their respective twelve-test slices at higher costs and twelve-month blocks, but no finding survives adjustment across the full 144-comparison grid. Tie-priority analyses show that some exploratory higher-cost outcomes depend on how exact score ties are resolved.Source: Table 10 · p. 15, §4.4 · p. 15, Figure 5 · p. 16
  • No factor-controlled partial-correlation test is significant after Holm correction; four Luna and seven Terra CSI 500 months have constant scores.Those constant-score months leave 68 Luna and 65 Terra identifiable months; other pool-model combinations have 72.Source: Table 11 · p. 15, §4.4 · p. 16

4.5 Input-intervention return results

  • All 24 paired intervention-return contrasts lack bootstrap-adjusted findings at six-month blocks; Luna’s row-order result depends on HAC adjustment and family size.Luna’s row-order effect is 0.542 percentage points per month: HAC-adjusted p=0.010 within three comparisons, versus 0.0815 in a pooled 24-test family; bootstrap-adjusted p=0.098. The three-model batch’s primary-condition mean CAGRs are Luna 7.29%, Terra 9.15%, and Sol 8.79%.Source: §4.5 · pp. 16–17

5 Discussion

The audit treats rankings as executable policies, combining repeated requests, frozen numerical rules, and matched portfolios to examine policy stability and returns. Rules approximate aggregate ranking responses more strongly than individual selection changes, while return findings and scope remain qualified by the tests and data used.

5 Discussion

  • Repeated requests measure decision variation, frozen rules test reproducible numerical structure, and matched portfolios measure returns associated with preferences beyond those rules.The paper frames a stock-ranking model as part of an executable economic policy.Source: §5 · p. 17
  • Frozen rules transfer to larger stock pools and reduce aggregate rank-change prediction error in feature-group removals and matched single-feature shifts, but event Jaccard of 0.000-0.125 indicates weaker prediction of individual entries and exits.The reported adjustments use separate families for the feature-group and single-feature comparisons.Source: §5 · p. 17
  • The primary 10-bp-per-side, six-month-block tests find no Holm-adjusted model-minus-rule return difference in the twelve-comparison family, and factor-controlled associations are also nonsignificant.Higher-cost Terra signals occur only in twelve-month-block sensitivity slices and do not survive adjustment across the full 144-test grid; the benchmark test does not establish formal equivalence.Source: §5 · p. 17

5.1 Scope and interpretation

  • The evidence covers observable policy under nine structured inputs, declared rule families, and a six-year Chinese equity sample, not internal reasoning or systems using news, retrieval, or external tools.The three universes share market months, and frequent CSI 500 boundary ties make outcomes sensitive to the specified anonymous-identifier tie rule.Source: §5.1 · p. 17
  • Retrospective model requests and market-data collection identify historical policy behavior rather than live forecasting ability.Service labels do not independently verify training cutoffs or backend snapshots; the matched-response design and final inference rule were fixed before collection.Source: §5.1 · p. 17

6 Conclusion

The section examines how development-fitted linear rules approximate models’ aggregate stock rankings, how return conclusions vary with evaluation choices, and how archived inputs and specified interventions structure the tests. Aggregate ranking approximation is distinct from less precise individual entry and exit predictions, while live forecasting remains for prospective evaluation.

6 Conclusion

  • Rules reduce aggregate rank-change error versus zero-change predictions in 12 archived feature-group and 8 matched single-feature comparisons, while individual entry and exit predictions remain less precise.The comparisons use the reported separate adjustment families.Source: §6 · p. 17
  • At 10 bp per side and six-month blocks, the 12-comparison model-minus-rule return family and factor-controlled association tests have no adjusted finding.Higher-cost Terra signals arise in 12-month-block sensitivity slices but do not survive adjustment over the full cost-by-block-length grid; the Luna row-order result depends on uncertainty procedure and comparison family.Source: §6 · p. 17
  • Frequent CSI 500 score ties make some simulated returns sensitive to tie priority, and live forecasting performance remains for prospective evaluation.Source: §6 · p. 17

S1 Experimental design and inference families

  • All historical comparisons are exploratory, and the primary comparison varies the ranking rule while holding the information set and portfolio engine fixed.The Luna, Terra and Sol historical grid contains 2,772 tasks; Astra has 924 under the same design.Source: S1 Experimental design and inference… · pp. 21–22

S2 Exact model inputs and scoring instructions

  • Models received numerical cross-sectional characteristic percentiles and assigned 0-to-1 relative-preference scores for selecting the 10 highest-scoring candidates.The prompt states that scores express relative preference, not calibrated return forecasts.Source: S2 Exact model inputs and scoring… · p. 23
  • Each universe and month has an archived matrix, anonymous mapping, display order, and schema; factor removals set specified columns to 0.5000, while no_semantics changes definitions and labels and row_order changes display order.The primary condition also has a separate archived input for each development signal.Source: S2 Exact model inputs and scoring… · p. 26

S3 Conventional training and parameter selection

Conventional-model settings are selected using development-period validation, with training labels restricted to returns completed before each signal. The appendix also documents how corporate events, successor claims, and trading constraints enter return labels and strategy execution.

S3 Conventional training and parameter selection

  • Hyperparameters are selected using mean monthly rank IC over 2017–2019, with validation labels exiting by 20191130; historical test returns are not used for selection.Ridge and histogram-gradient-boosting candidate settings are enumerated in the text, and the parameter table is titled as covering six universe-condition configurations.Source: S3 Conventional training and parameter… · p. 26, Table S5 · p. 26
  • Validation folds train only on labels completed before their first signal, and candidate performance is the mean monthly Spearman correlation between predictions and stock returns.The expanding-sample table is identified as covering training samples; training audits retain all 72 historical test signals.Source: Table S6 · p. 26, S3 Conventional training and parameter… · p. 27
  • Stock-return labels track the economic claim from entry to exit, while strategy purchases additionally require an executable opening quote and compliance with price-limit checks.The common wealth calculation includes share exchanges, locked entitlements, and board transfers; claims on delisted securities may remain invested until a scheduled disposal is possible.Source: S3 Conventional training and parameter… · p. 27
  • Successor claims retain entitlement and valuation records, and event and board-price verification is bounded by the archived window ending 06 January 2026.Daily snapshots accompany each cost scenario; later settlement windows require a dated verified event and quotation version.Source: S3 Conventional training and parameter… · p. 27

S4 Data coverage and missing characteristics

The appendix documents panel coverage, feature-missingness accounting, and source-history checks for the stock universe. It also records the limits of the retained historical records and a specific identity crosswalk.

S4 Data coverage and missing characteristics

  • Each universe panel contains 132 monthly signals, including 60 development and 72 test signals, and the coverage table is organized around contemporaneous panel coverage.The passage notes that monthly membership, eligibility, and exclusion counts are recorded by universe.Source: Table S7 · p. 27, S4 Data coverage and missing… · p. 28
  • Missing-characteristic counts are calculated before within-month ranking among eligible stocks, with missing ranks assigned 0.5; these counts describe feature availability, not missing prediction outcomes.The missing-feature table is identified as reporting missing feature observations.Source: S4 Data coverage and missing… · p. 28, Table S8 · p. 28
  • The quote audit covers 1,551 stock-level price/adjustment pairs and 4,105,829 price rows, checking date coverage, exchange dates, adjustment alignment, and valid price and volume observations.The official Shenwan history covers 1,551 inventory stocks across 4,160 dated classification records; the analysis is a retrospective risk diagnostic.Source: S4 Data coverage and missing… · p. 28
  • A dated identity crosswalk maps old numeric code 300114 to canonical code 302132 on the Shenzhen exchange, effective 20250217.The mapping is supported by an issuer announcement dated 20250214, while linked intervals retain original classification fields alongside the canonical code.Source: S4 Data coverage and missing… · p. 28

S5 Historical return inference and intervention returns

The section specifies paired historical-return and input-intervention comparisons, alongside risk-matched and disjoint-candidate portfolio diagnostics. Risk matching separates target constraints from exposures observed after execution.

S5 Historical return inference and intervention returns

  • Historical return contrasts pair scheduled model trajectories over calendar months, with separate three-test model-proxy and twelve-test risk-matched families.Comparisons use 10 basis points per side; six-month block results are adjusted within declared families, not one global family.Source: S5 Historical return inference and… · pp. 29–30
  • Altered-input return entries subtract primary run 1; the first three columns also subtract the matching EqualFactor change.Each altered condition has one complete run, so these contrasts do not separately estimate request-to-request variance.Source: S5 Historical return inference and… · pp. 29, 31, Table S13 · p. 31

S6 Preference fits, decision stability and risk diagnostics

  • At 10 bp, three-model cohort maximum observed industry deviations were 5.319 percentage points (SSE50), 3.999 percentage points (CSI300), and 4.142 percentage points (CSI500).Astra's corresponding deviations were 2.406, 3.998, and 2.981 percentage points, respectively.Source: S6 Preference fits, decision stability… · p. 32
  • Risk matching maximizes portfolio-weighted ranking scores while constraining target weights and industry, size, volatility, and beta exposures around an equal-weight reference.Post-trade exposure readings are separate from target constraints and use actual 09:30 holdings after monthly rebalancing.Source: S6 Preference fits, decision stability… · p. 32
  • Risk variables use retrospective provider records; original release snapshots and revision histories are not reconstructed.The reported exposure diagnostics therefore do not reproduce historical provider vintages.Source: S6 Preference fits, decision stability… · p. 32
  • The disjoint-candidate experiment removes overlapping stocks across universes while retaining shared market-month shocks, comparing LLM and equal-factor portfolios.Its twelve contrasts use 10 basis points per side.Source: S6 Preference fits, decision stability… · p. 33

S7 Development-based return power and precision

Development-based planning selects a 36-month duration, while the paper separately reports historical prediction coverage and frozen-rule agreement across universes. The selected duration is a planning output, not a change to the historical sample.

S7 Development-based return power and precision

  • The duration calculation uses one 60-month development trajectory per model and universe, with HAC lag-3 variance and a 0.10-percentage-point monthly effect.The three CSI 300 comparisons determine the archived duration calculation; CSI 500 supplies additional variance estimates.Source: S7 Development-based return power and… · p. 33
  • 36 months is selected because the three CSI 300 comparisons do not all reach 80% planned power; model-specific planning power is 2.84%-4.11%.The calculation uses development-period returns and is separate from the observed 72-month historical sample.Source: S7 Development-based return power and… · pp. 33–34, Table S17 · p. 33

S8 Model configuration and completed prediction coverage

  • The historical grid contains 3,696 valid tasks, including 924 Astra tasks; the separate return-planning batch covers Luna, Terra, and Sol.The matched-response grid is excluded from historical coverage and contributes 216 responses per model.Source: S8 Model configuration and completed… · pp. 34–35

S9 Frozen-rule agreement and annual transfer

  • Annual results are descriptive, and holding years refer to portfolio entry rather than the signal date.Disjoint-subset rankings are recomputed within retained stocks from stored full-pool responses, without new model calls.Source: S9 Frozen-rule agreement and annual… · pp. 35, 38

S10 Rule-complexity inference

The section compares bounded-forest and frozen-linear rule agreement, then reports archived response-prediction and directional diagnostics. MSE gains are reported within separate historical and Astra comparison families.

S10 Rule-complexity inference

  • Rule-complexity differences subtract frozen-linear rank agreement from bounded-forest rank agreement, with three models and Astra adjusted as separate families.Intervals are pointwise block-bootstrap intervals at each declared block length.Source: S10 Rule-complexity inference · p. 39, Table S23 · p. 39

S11 Archived response prediction and directional diagnostics

S12 Matched single-feature experiment

The matched experiment changes one numerical feature for selected stocks while holding other prompt inputs fixed. It measures response changes and portfolio-return effects using repeated responses within a month.

S12 Matched single-feature experiment

  • The matched grid includes 24 months, four models, three conditions, and three responses per condition, totaling 864 valid responses.The conditions are original input, valuation shift, and momentum shift; Astra contributes 216 responses as a replication batch.Source: S12 Matched single-feature experiment · pp. 41–42, Table S27 · p. 41
  • Endpoints include signed response, direction agreement, selection overlap, and event Jaccard; event Jaccard is conditional on a nonempty event union.These are descriptive diagnostics of response and discrete selection changes.Source: S12 Matched single-feature experiment · p. 42
  • Eight stocks per month receive signed 0.2 shifts on book-to-price or 12-to-1-month momentum, with the same stocks and signs for both features.Other numerical cells, identifiers, row order, and prompt components are held fixed; these are numerical input changes, not simulated fundamental changes.Source: S12 Matched single-feature experiment · p. 42, Table S29 · p. 42
  • Responses are averaged within condition before rank changes are computed, and inference resamples month clusters within holding-year strata.The predeclared precision target concerns conditional measurement noise in signed rank response, not financial-return power or total treatment-response heterogeneity.Source: S12 Matched single-feature experiment · pp. 41–42

S13 Incremental investment value and remaining-information sensitivity

The section reports net-return and remaining-information sensitivity analyses with Holm adjustments within comparison slices. Across the full cost-by-block-length grid, the minimum adjusted p-value is 0.101.

S13 Incremental investment value and remaining-information sensitivity

S14 Astra input-intervention results

The Astra intervention study compares altered conditions with a within-batch primary condition across six conditions and three runs. Its return contrasts use 10 basis points per side and six-month-block intervals.

S14 Astra input-intervention results

S15 Luna, Terra and Sol input-intervention results

This section reports performance and score-rank results from input interventions spanning 3,888 valid responses across 18 model-condition groups, and documents the associated data and reproduction paths. No six-month-block bootstrap test is significant after adjustment; Luna’s row-order result varies with the uncertainty method and comparison family.

S15 Luna, Terra and Sol input-intervention results

  • The intervention design contains 3,888 valid responses across 18 model-condition groups, with 216 responses per group.The 432 input matrices implement six conditions; all 54 trajectories use the same trading engine and 72-month calendar, with costs of 0, 10, 25, and 50 basis points per side.Source: S15 Luna, Terra and Sol… · p. 50
  • All 24 six-month-block bootstrap tests have adjusted p-values above 0.05.For Luna, reordering rows changes mean monthly net return by 0.542 percentage points (pointwise interval [0.154, 1.125]); the bootstrap Holm-adjusted p-value is 0.098, versus 0.010 under HAC within the three row-order comparisons, but 0.0815 across all 24 contrasts. The interval is pointwise.Source: S15 Luna, Terra and Sol… · p. 50, Table S34 · p. 51
  • The 24 six-month contrasts comprise factor-neutralization, semantic, row-order, and EqualFactor-adjusted factor comparisons, tested in separate 9/3/3/9 Holm families.Inference uses three whole paired trajectories and 72 shared months; HAC uses three lags, and intervals are pointwise.Source: S15 Luna, Terra and Sol… · p. 52

Input-intervention results

  • Table S39 presents three-run performance distributions at 50 bp, with CAGR and total return compounded within each run before summarization.Monthly return is the arithmetic mean of 72 holding-period returns; volatility is annualized, Sharpe uses a zero reference rate, and drawdown uses daily NAV.Source: Table S39 · p. 57
  • Table S40 reports descriptive score-rank agreement across repeats and interventions, averaging defined pair correlations within months and then across defined monthly means.Constant-score correlations are undefined, and the primary condition has no intervention comparison.Source: Table S40 · p. 58, S15 Luna, Terra and Sol… · p. 58
  • Table S41 identifies the input-intervention data and reproduction paths; portfolio reproduction requires authorized SSE 50 market data and related local inputs.The release includes monthly returns, NAV, and aggregate execution records; stock-level trades and terminal positions remain in the local study archive.Source: Table S41 · p. 58, S15 Luna, Terra and Sol… · p. 59

S16 Reproduction and source locations

The reproduction materials preserve response-level records and code, while portfolio analyses require separately authorized market sources. The section also documents analysis commands, validation, and manuscript rendering.

S16 Reproduction and source locations

  • The research archive preserves 1,260 input files, 10,104 unique valid responses, fitted rules, rankings, and derived portfolio statistics.Portfolio recomputation also requires separately authorized materials; source permissions constrain redistribution of some retained local files.Source: S16 Reproduction and source locations · pp. 59–60
  • Portfolio reproduction requires a separate research directory with licensed source archives and empty derived-output locations.The reproduction sequence prepares analyses for SSE 50, CSI 300, and CSI 500.Source: S16 Reproduction and source locations · p. 60
  • The documented analysis sequence includes panel and baseline preparation, main analysis, mechanisms, surrogates, statistics, and validation stages.The three universe-specific analyses support joint historical comparisons, planning, and final validation.Source: S16 Reproduction and source locations · p. 60
  • The public archive includes 34 test files, but complete test discovery also requires local authorized source panels.Source: S16 Reproduction and source locations · p. 60
  • Response-based reproduction uses stored model answers, development-fitted coefficients, and controlled inputs; portfolio analyses use authorized research data.Source: S16 Reproduction and source locations · p. 60
  • The manuscript renderer builds journal and arXiv editions from one content model and requires a local template and separately supplied Tectonic executable.Portfolio recomputation additionally requires authorized observations and economic-event and board-price records.Source: S16 Reproduction and source locations · p. 61
  • Astra analysis, rule, economics, and controls scripts reproduce its intervention analyses; companion scripts format tables and generate five figures.Source: S16 Reproduction and source locations · p. 61
  • Versioned inputs, code, and output hashes identify analyses, while portfolio records reconcile costs, entitlements, daily values, and terminal wealth.The audit covers 2,436 Astra responses and 3,888 responses in the three-model replication.Source: S16 Reproduction and source locations · p. 61

S17 Selection-boundary tie sensitivity

Tie-priority sensitivity tests hold model scores and candidate sets fixed while varying the ordering of exactly tied scores. The primary return finding is unchanged, while exploratory higher-cost results vary with the tie rule.

S17 Selection-boundary tie sensitivity

  • The analysis varies priority only within equal-score groups, using 50 seeded priorities and ascending anonymous identifiers as the reference.Each priority uses the same self-financing engine, and reference-priority monthly returns agree with stored returns to within 1e-12.Source: S17 Selection-boundary tie sensitivity · pp. 61–62
  • Boundary ties are common in CSI 300 and CSI 500, especially for Luna and Terra in CSI 500; whole-cross-section constant scores occur in four Luna and seven Terra months.Table S43 summarizes selection-boundary ties and 10-basis-point return sensitivity.Source: S17 Selection-boundary tie sensitivity · p. 61, Table S43 · p. 62
  • At 10 basis points per side and six-month blocks, no alternative priority yields an adjusted p-value below 0.05 in the twelve-comparison family.The comparison uses the same score vectors and market months; alternative priorities do not add independent market observations.Source: Table S44 · p. 62, S17 Selection-boundary tie sensitivity · p. 62
  • At higher costs with twelve-month blocks, Terra findings vary across priorities: the reference has three, while alternative-priority counts are 26, 50, and 40 cases.CSI 500 Terra at 25 basis points crosses the threshold under 17 priorities; the full four-cost-by-three-block-length grid has no adjusted finding across 144 comparisons.Source: S17 Selection-boundary tie sensitivity · pp. 62–63, Table S45 · p. 63
Loading 2610.01213v1…