Earnings call participation network and dynamics as an alternative data factor for US equities.

01Participation as a source of information

An earnings call records what a company reports and which external analysts ask questions. This study asks whether the history of that participation can improve stock selection beyond a model built from prices and trading volumes.

External analysts participate in a professional research capacity. Their work brings sector knowledge, familiarity with company reporting and comparisons across businesses into the discussion. A question may reflect extensive analysis, expectations and unresolved concerns formed before the call. Analysts differ in experience, expertise and access, so the information associated with their participation need not be equivalent.

When the same analysts recur across companies and reporting periods, their participation forms an observable network. Its changing connections provide a way to study how professional attention is distributed and how its focus shifts. ECPND converts this history into a daily research factor. Our hypothesis is that participation patterns contain information relevant to subsequent returns that is not fully captured by a model built from prices and trading volumes.

What the research suggests

Earlier work establishes two relevant connections. Mayew, Sharp and Venkatachalam (2013) find that analysts asking questions issue more accurate and timely annual earnings forecasts immediately after calls than nonparticipants [1]. Mayew (2008) finds that analysts with more favourable stock recommendations are more likely to obtain access to questioning [2]. Participation can therefore reflect both research activity and the opportunities management provides to engage.

Research published in the 2020s extends this picture. Bradley, Gokkaya and Liu (2020) associate professional connections between analysts and executives with better forecast accuracy, more informative recommendations and greater participation in calls [3]. Rennekamp, Sethuraman and Steenhoven (2022) examine the interactions themselves: greater conversational engagement, measured through linguistic style matching, is associated with larger absolute price movements during the discussion [4]. This supports studying interaction quality alongside participant identity.

Two studies published in 2026 address the significance of analyst expertise more directly. Awyong, Cho and Yang find that participation is associated with subsequent revisions to management forecasts, with stronger associations when the analysts are more experienced and have greater forecasting ability [5]. Holowczak, Jiu, Kyung and Yu link all-star analyst participation to abnormal returns during the Q&A session, with a stronger relationship for analysts whose forecasts are more accurate [6]. These findings give an empirical basis for considering who participates, as well as how many people participate.

The studies examine different measures and outcomes from ECPND. They motivate its economic rationale; the score’s contribution to future portfolio returns requires its own test. Company size, sector conditions, coverage practices and access to management may also shape the observed network. Our paired experiment holds forecasts, capital, execution rules and costs constant, then measures what changes when ECPND is used in stock selection. Separating the economic mechanisms behind any contribution remains a further research question.

03From participation to ECPND

The score summarizes eligible participation history as a daily measure, indexed by date and ticker for use alongside model forecasts in systematic equity research.

From source observation to portfolio comparison

1. Observe. Read the issuer, call date and named external participants from structured transcript components.
2. Connect. Normalize recurring identities and connect eligible participation observations across companies.
3. Score. Evaluate the eligible history and publish a bounded daily value, retaining missing observations explicitly.
4. Evaluate. Use the score in a fixed selection overlay and compare the resulting portfolio with the unchanged baseline.

Participation refers to named external analysts whose spoken contributions are recorded in the transcript. Management and issuer representatives are excluded. Recurring identities link participation across companies and over time. The score evaluates this history using outcomes from other companies, with return periods completed before the scoring date. The issuer’s own outcomes are excluded from these comparisons, and estimates based on limited histories receive less influence.

• Field: ecpnd_score, a daily value between 0 and 1.
• Direction: Higher values are more favourable; neither a probability nor a return forecast.
• Persistence: Daily scores combine the latest eligible call with evolving network evidence.
• Missing score: No eligible value for that stock and date: blank in CSV, null in the API.

Daily scores and network evolution

ECPND evaluates each company through its connections to the wider earnings-call participation network. The company’s latest eligible call identifies the participants relevant to its score, while their eligible history across other companies provides the broader context. New calls refresh participation records, and completed outcomes from earlier calls update the evidence associated with shared participants. These changes can affect several connected companies, allowing scores to evolve even when an individual company has no new call.

The freshness limit applies to the company’s own latest eligible call record, which expires at an age of 126 US trading sessions, roughly six months. Age starts at zero on the first session after the call date; recalculation does not reset it. A newer eligible call replaces the earlier record, subject to the same participation and history requirements. If the current record expires, the daily score becomes missing.

In the historical score file, date identifies the session to which the value applies; use the supplied score for each portfolio decision session. For live decisions, use a snapshot published before the decision and check its update times and data status. Buy & hold retains its initial positions despite later score changes or expiry. Weekly rebalancing uses the current scores and stated missing-score policy at each scheduled decision.

What does “missing score” mean?

A missing score describes the availability of a measurement. It means the inputs eligible under the declared timing convention do not support an ECPND value for that stock on that date. It is not a negative opinion about the company, a neutral score, or a forecast of poor returns. A numerical score of 0 is a valid value and is different from a blank.

• No eligible prior call: No source call is eligible for that stock and date.
• No usable participants: The latest eligible call has no identifiable external analysts.
• Insufficient history: Eligible peer history does not meet the evidence requirements.
• Outdated call record: The latest record has reached the 126-session expiry limit.

Scores between calls. An eligible participation record can support daily scores between a company’s calls. The score uses the evidence available for each date, including newly completed historical outcomes linked through shared participants at other companies. It can therefore evolve without a new call from the company. For portfolio selection, use the score supplied for the decision date. Preserve missing values; substituting an earlier score or zero would change the stated selection rule.

Treatment in the portfolio tests. In the primary study, stocks without an ECPND score remain eligible for selection based on the baseline model’s ranking. They receive no factor penalty and are excluded from the calculation of score percentiles. The scored-only control removes these stocks from both portfolio arms before selection, allowing the overlay to be evaluated within a common universe of scored stocks.

Scope of disclosure. This paper documents the factor’s economic rationale, data inputs, timing conventions, study design and results. The detailed methods for resolving participant identities and converting participation history into ECPND scores remain proprietary. Section 4.2 specifies how the supplied scores are applied to portfolio selection, including percentile ranking, the selection threshold and the treatment of missing values. Researchers can use these rules to evaluate the delivered factor without reproducing its underlying calculation.

04The paired portfolio design

Every comparison begins with the same model forecasts. ECPND changes the stock-selection decision; the capital, execution convention and accounting rules are shared.

4.1 Baseline model and training chronology

The baseline combines Microsoft’s Qlib research framework [7] with LightGBM, a gradient-boosted decision-tree model [8]. It uses 157 price and volume features from the Alpha158 family [9] to rank eligible stocks; the VWAP feature is omitted because that input is unavailable in the study data. The training target is the adjusted return over the next 42 trading sessions, measured from the first session’s open to the last session’s close. These returns are standardized across stocks on each training date.

Models are fitted annually on expanding historical windows. Every training return must have ended before the test year begins. The run uses 200 boosting rounds and four fixed random seeds - 19, 41, 73 and 101 - to expose variation from randomized training choices. The seeds share the same underlying market history. Both portfolio arms reuse the frozen predictions; the baseline is not retrained with ECPND.

Role of the published software. Qlib and LightGBM provide the modelling foundation. Our US data preparation and weekly portfolio simulator are custom implementations, rather than an unchanged official Qlib benchmark or its daily TopkDropout strategy. ECPND has no implied endorsement from the software authors.

4.2 Which score changes the selection?

The factor input is the daily ecpnd_score available for each stock at the portfolio decision. It has a different job from the Qlib / LightGBM prediction: the model prediction ranks stocks, while ECPND adjusts their selection priority.

The primary study gives lower selection priority to the lowest 20% of available scores in the eligible universe on each decision date. This is a relative rank threshold, not a raw ECPND value of 0.20. The application is specified as follows.

1. Use the score for the execution date. Join by ticker and CSV date, using the same eligible universe and prior-close model predictions for both arms. The CSV date already reflects score eligibility at that session’s open; do not add another day of delay, use later observations or fill unavailable rows with older scores.
2. Rank the available scores before selecting holdings. Sort scores from lowest to highest across the entire eligible universe for that date. Divide each stock’s average ordinal rank by the number of scored stocks, N. Ties receive the average of their occupied rank positions. Deprioritize percentile ranks ≤ 0.20 when N ≥ 30. With fewer scored stocks, keep the baseline ranking. Missing scores are excluded from this calculation and receive no penalty; a numerical zero remains a valid score.
3. Select 50 stocks with a deterministic order. The baseline sorts model predictions from highest to lowest, breaking ties alphabetically by ticker. The factor arm puts unpenalized stocks first, then penalized stocks, using the same model-prediction and ticker order within each group. Take the first 50. Penalized stocks can fill remaining places when fewer than 50 unpenalized candidates exist. With fewer than 50 eligible stocks, no portfolio is formed.
4. Apply the rule at the scheduled decision. Buy & hold uses it at entry and retains the share quantities. Weekly rebalance repeats it at each weekly decision with current eligible scores and predictions, and resets target weights under the shared allocation and cost rules in Section 4.3.

Example. Among 100 scored eligible stocks without ties, ranks 1–20 receive lower priority. The corresponding raw-score boundary changes with the date and universe. Ties may change the penalized share. A high score alone does not guarantee selection: the model ranking still determines which stocks fill the 50 places.

• Baseline: Select the 50 highest-ranked eligible stocks using the unchanged Qlib / LightGBM forecasts.
• With ECPND: Adjust the same model ranking using ECPND, then select 50 stocks.

The baseline starts with price and volume information available at the preceding market close. Qlib computes the features, and the annually trained LightGBM model produces forecasts that rank eligible stocks. Both portfolios use these same forecasts; ECPND adjusts selection priority under the rule above. Capital, weighting, execution and trading costs follow the same rules in both arms. Their return difference measures the overlay’s contribution within this specific model and portfolio design.

4.3 Portfolio and execution assumptions

• Formation: Each week, a new baseline/factor pair starts at the first eligible US trading session, normally Monday. Every account begins with $1 million and is tracked separately from earlier weekly formations.
• Allocation: Allocate 95% of pre-trade equity equally across 50 positions at entry or rebalance. Fractional adjusted shares permit exact allocations; fees reduce the cash reserve, which earns no interest.
• Buy & hold: Retain the initial share quantities throughout each evaluation horizon. Later score changes or expiry do not trigger trades, so weights can drift as prices change and no scheduled rebalancing takes place.
• Weekly rebalance: Reapply each arm’s selection rule at the first US trading session of every week. Reset equal target weights using current account equity, including adjustments to retained positions as well as stock replacements.
• Trading costs: Charge 10 bp (0.10%) of every executed buy or sell notional, including trades that resize retained positions. Both arms use the same fee rule; spread, market impact and capacity are not separately modelled.
• Prices: Execute trades at frozen adjusted opening prices and value each account daily at adjusted closing prices plus cash. Corporate actions are reflected in the adjusted series, with no additional dividend cash flow.
• Evaluation: Evaluate returns after 3, 6 and 12 calendar months, using the last session close on or before each anniversary. Positions remain open at the valuation date, so no forced liquidation or terminal sale fee is applied.

Each formation produces 16 accounts: four seeds × two portfolio policies × two arms. The study therefore comprises 3,920 portfolio records across 245 formation weeks. Stored horizon records cover 3 and 6 months; a separate revalidation extends the analysis to 12 months.

4.4 Return difference and uncertainty

For each matched portfolio pair, we subtract the baseline return from the return with ECPND, after the stated trading costs. We average the four seed differences within each starting week, then average those weekly results across all completed formations. Each starting week carries the same weight. The calculation is repeated separately for buy & hold and weekly rebalancing at each holding horizon.

One basis point is 0.01 percentage points. A six-month difference of +54 bp means that the factor portfolio returned 0.54 percentage points more over that holding period. It is a net return difference under the stated costs, without annualization or adjustment for other equity-factor exposures.

Overlapping weekly portfolios share much of their holding period. Treating them as independent tests would overstate the amount of evidence. We therefore resample contiguous weekly blocks, with circular wrapping, to estimate uncertainty in the mean [10].

• Resamples: 5,000.
• Block lengths: 13 / 26 / 52 weeks.
• Reported interval: 2.5th to 97.5th percentiles.

Block lengths follow the 3-, 6- and 12-month horizons. The resulting intervals describe uncertainty under this resampling design. They do not account for every prior research choice, and long blocks leave relatively few distinct historical periods to resample.

05Incremental portfolio returns

The mean difference is positive in all six primary comparisons. Evidence is strongest for the six-month buy & hold result, while the remaining primary intervals include zero.

For buy & hold, mean uplift is +17, +54 and +107 bp at 3, 6 and 12 months. Weekly rebalancing produces +28, +52 and +97 bp. These are averages of completed, matched formations after the primary trading-cost assumption (see Appendix A.1).

Figure 1Mean uplift and its uncertainty

Reading the table. Baseline and factor returns are means at the stated horizon. “Positive” is the share of formation weeks with a positive difference after averaging seeds. Formation counts are starting dates, not independent market histories; maturity determines which dates enter each horizon.

5.1 Variation across formation years

The year groups show six-month portfolio outcomes by starting year, rather than calendar-year strategy returns. Both policies underperform for portfolios formed in 2023. Other years contribute positive differences, while 2026 is mixed and includes only 11 completed formations. The factor’s historical contribution is therefore sensitive to the period being evaluated.

Figure 2Six-month uplift by formation year

5.2 Outlook: improving consistency across configurations

Further research examines how selection strength and complementary information affect the size and consistency of the contribution. The threshold experiments show that these need separate consideration. Deprioritizing the lowest 30% rather than 20% of ECPND scores raises mean six-month buy & hold uplift from +54 to +89 bp. The 95% interval is +13 to +99 bp under the 20% rule and +33 to +151 bp under the 30% rule. The estimated contribution increases, while the uncertainty range becomes wider.

The combined-factor study examines ECQDI (Earnings Call Quantitative Dynamics and Intensity) as a complementary input. Weighting ECPND ranks at 75% and ECQDI ranks at 25%, while retaining the 20% lower-priority threshold, produces +69 bp of six-month buy & hold uplift versus Qlib. Its 95% interval of +41 to +99 bp is narrower than ECPND's primary interval. The two scores have a low average cross-sectional rank correlation of +0.076, giving a concrete reason to investigate their complementary use. The narrower interval concerns uncertainty in the historical mean; it does not establish lower portfolio risk.

These findings identify promising configurations for further evaluation. The combination's incremental interval versus ECPND alone still includes zero; later-period and continuous-account checks do not establish a consistent advantage. The intervals are not adjusted for comparing multiple settings on this previously researched history. The next stage is to fix a small set of thresholds and weights before evaluating new observations, alongside controls for sector, size, momentum and execution costs. The objective is a contribution that remains useful across market conditions and independently chosen portfolios.

06Testing the assumptions

Does the result depend on our chosen costs, selection threshold or stock coverage? We change one assumption at a time and repeat the six-month comparison, keeping the same model forecasts.

6.1 What happens when trading costs rise?

Every buy and sell incurs a cost. We test 0, 10, 25 and 50 bp of the amount traded, charging both portfolios at the same rate.

Buy & hold: The factor adds +54 bp at all four cost levels. Both portfolios invest the same amount at entry and make no subsequent trades. Higher fees therefore reduce their returns equally.

Weekly rebalance: Costs accumulate with repeated trading. At 10 bp, mean returns are 5.34% for the baseline and 5.86% with ECPND. At 50 bp, those returns are -7.18% and -6.35%. The factor portfolio trades less on average, which reduces the impact of fees, but both portfolios lose money on average at the highest tested cost.

This check applies fixed trading costs. Stock-specific spreads, market impact and trading capacity require separate evaluation.

6.2 Does a different score threshold help?

We give lower priority to the lowest 10%, 20% or 30% of available scores. At the standard 10 bp trading cost, moving from 20% to 30% raises buy & hold uplift from +54 to +89 bp, while weekly-rebalance uplift falls from +52 to +38 bp. A stronger score filter can help one portfolio approach and weaken another. The primary study retains the 20% rule; the wider threshold study examines further settings.

6.3 Does the stock coverage explain the gain?

Two further checks keep the primary selection and cost rules, but change which stocks can enter either portfolio.

• Only stocks with scores: Remove unscored stocks from both portfolios before selection. ECPND still adds +63 bp for buy & hold and +50 bp for weekly rebalance. The positive mean contribution therefore does not require holding unscored stocks.
• Check unusual price moves: Remove 15 tickers associated with 88 extreme overnight gaps. The remaining stocks produce +61 bp and +50 bp, respectively. The mean contribution remains positive after these tickers are removed.

The price check identifies unusual moves using the full historical record, including later dates. It helps assess the influence of those tickers; it is not a stock-selection rule available in advance.

These checks reuse the same historical period. They show how the result changes under alternative assumptions; new observations are still needed to test whether it persists.

07Continuously invested accounts

A series of overlapping portfolio cohorts answers a different question from the experience of a single account carried through time. We report both.

For each seed and arm, one account starts on 10 January 2022 and is rebalanced weekly through 18 September 2026, with the primary 10 bp cost assumption. No fresh capital is introduced for subsequent formation weeks.

Figure 3Annualized growth across the four model seeds

The chart follows the same starting capital through repeated weekly decisions. For seed 19, the simulated account grows at an annualized 6.40% under the baseline and 8.65% with ECPND, after the stated trading costs. Each week’s gains or losses carry into the next allocation. The comparison therefore captures the cumulative effect of stock selection, rebalancing and compounding over the full period.

The factor improves three of the four seeds and slightly trails the baseline in seed 101. The mean of the four paired CAGR differences is +1.09 percentage points per year. This is an average across model runs, rather than the CAGR of a simulated blended portfolio. The differences range from slightly negative to materially positive, showing that the factor’s contribution depends in part on the baseline ranking.

• CAGR: Annual growth rate equivalent to the account’s total return.
• Maximum drawdown: Largest peak-to-trough decline in daily account value.
• Sharpe ratio: Daily mean return / volatility, annualized over 252 sessions; no risk-free adjustment.
• Information ratio: Mean / volatility of daily excess returns vs. baseline, annualized (252 sessions).

08What the evidence supports

ECPND improves average returns in the tested Qlib / LightGBM portfolios. These results support evaluating participation history as an additional input to US equity selection.

The clearest primary evidence comes from six-month buy & hold, where the 95% interval lies above zero. In the other five primary comparisons, average gains are positive but the intervals also allow for no improvement or underperformance. Results vary across starting years and model seeds. The practical question is therefore how consistently ECPND adds value across portfolio rules, market conditions and trading costs.

Five areas define the next stage of validation:

• Information timing: The next-session rule leaves a buffer after same-day transcript delivery. Live snapshots record publication times; historical transcript versions and later revisions require separate verification.
• Universe coverage: Both arms share the same eligible universe. Historical membership, delisted stocks and adjusted-price coverage need fuller verification to assess their effect on the comparison.
• Research selection: Training uses completed historical outcomes. Comparing thresholds and variants on the same history adds selection uncertainty that the reported intervals do not account for.
• Portfolio exposures: Sector, size, momentum and liquidity are not fully controlled. Some uplift may reflect these exposures, so the return difference is not an estimate of unexplained, risk-adjusted alpha.
• Live execution: Fixed fees do not fully represent spreads, market impact or trading at a given order size. Publication latency and performance on newly arriving data require strategy-specific evaluation.

Forward evaluation should fix the score version and portfolio rules before new outcomes arrive, save each delivered snapshot with its publication time, and retain corrections as separate records. Paired portfolios can then track the factor’s contribution under shared execution and cost assumptions. Tests in independently chosen models, with portfolio exposures controlled, would help establish how broadly the historical result carries into practice.

09Verification and research access

The research records link the reported averages to portfolio accounts, trading assumptions and downloadable score observations.

9.1 Internal consistency checks

Internal checks address two questions: whether the reported return differences can be reproduced from the frozen study inputs, and whether recorded trades reconcile to daily portfolio values. Appendix A documents the dated runs, their scope and their results. These checks assess computational consistency; they do not constitute an external audit or verify historical transcript publication times.

9.2 The evaluation procedure

The evaluation uses supplied ECPND scores. Follow points 1-8 for each eligible weekly formation and each of the four fixed model seeds, then summarize the paired results in points 9-10. Keep each policy and horizon separate.

1. Load inputs. Join frozen forecasts and eligible scores for the decision date.
2. Select the baseline. Take the top 50 eligible stocks by unchanged model rank.
3. Rank scores. Deprioritize the lowest 20% when at least 30 stocks are scored.
4. Select with ECPND. Take 50 stocks using the priority, tie and missing-score rules in 4.2.
5. Match assumptions. Use identical capital, allocation, execution and trading costs.
6. Run both policies. Simulate buy & hold and weekly rebalance separately.
7. Value each horizon. Use the last close on or before each completed 3-, 6- or 12-month horizon.
8. Calculate the difference. Subtract baseline return from factor return.
9. Average seeds. Average the four paired differences within each formation.
10. Summarize formations. Average completed weeks equally; resample week blocks for uncertainty.

9.3 Data for your own evaluation

The public package starts with five numbered ECPND files: 00_README_ecpnd.md, the daily score CSV, study results, study cohorts and a worked selection example. The score sample covers 1 August 2024–31 July 2026, a two-year window with an eight-week delay. It contains 244,757 observations across 523 tickers and 501 market sessions, including missing scores.

The folder 05_additional_research_ecqdi introduces ECQDI - Earnings Call Quantitative Dynamics & Intensity, a separate text factor derived from external analysts’ questions. It applies a fixed, deterministic measurement to their quantitative content and ranks each call against earlier calls in the same sector. The rationale is that quantitative questioning may reflect the estimates, assumptions and financial considerations analysts bring to the discussion. ECPND studies participation history and network relationships; ECQDI examines the quantitative character of the questions being asked.

The supplied field, ecqdi_score, ranges from 0 to 1, with higher values indicating greater quantitative intensity relative to its historical sector reference. The companion pack follows the same five-file format: a short README, two years of daily scores, study results, cohort comparisons and a worked selection example. Both score samples cover the same dates and eligible tickers, allowing evaluation of ECQDI on its own or alongside ECPND. Its score interpretation, eligibility rules and results are explained in the ECQDI research note. It is provided as additional research, with separate files and results from the ECPND product.

• date: Session date with score eligibility at the open under the study’s next-session rule.
• ticker: Stock identifier in the eligible research universe.
• ecpnd_score: Value from 0 to 1; blank means no eligible score for that stock and date.

The selection example follows one historical formation through the primary 20% rule. Each row shows the stock, baseline prediction and rank, ECPND score and percentile, selection priority, final rank and candidate = yes/no. All eligible stocks are shown, so a reader can see both stocks leaving the baseline selection and replacements entering from below its original cutoff. The README explains the column order and the decision reasons.

The sample supports a new evaluation over its own dates. The full study begins earlier, in 2022; its earlier scores, full baseline forecast history and market prices are outside the download. One day of seed-19 forecasts is supplied for the worked example. The selection rule is disclosed, but an exact replay of the whole study also requires the remaining inputs. Use your own forecasts and prices to evaluate ECPND in a new, explicitly defined experiment.

10Outlook and further research

The follow-up asks how strongly ECPND should influence stock selection, and whether its positive mean contribution remains when both portfolios hold only stocks with available scores.

Two completed extensions address these questions. The first varies the share of scores receiving lower selection priority. The second repeats that comparison after removing stocks without a score from both portfolios. Together, they examine the application of ECPND while keeping its calculation unchanged.

10.1 Giving the score more influence

We test nine buy & hold settings, giving lower priority to the lowest 10% through 90% of available scores in ten-percentage-point steps. Each uses the same frozen forecasts, scores and prices, four model seeds, 50-stock allocation, weekly formation dates and 10 bp trading costs. Timing and holding periods follow Section 4; Appendix A.3 documents the extension.

A 50% setting moves the lower half of scored candidates behind the remaining candidates. Model predictions still determine the order within each group. The threshold refers to a percentile rank, rather than a raw score of 0.50. The minimum score count, tie handling and missing-score treatment remain those in Section 4.2: unscored stocks receive no penalty, and lower-priority stocks can fill places if needed.

Completed formations: 232 at 3 months, 219 at 6 months, 193 at 12 months. Each formation averages four seeds. Returns are after 10 bp trading costs, without annualization. Intervals and the average number of entry holdings without a score refer to the six-month sample.

10.2 Higher averages, different portfolios

The six-month mean difference rises from +54 bp at the primary 20% setting to +89 bp at 30% and +105 bp at 50%. The 80% and 90% settings reach +232 and +239 bp. The average generally increases as ECPND receives more influence, although the progression is not uniform.

Higher averages come with less precise estimates. The six-month 95% interval widens from [+13, +99] bp at 20% to [-107, +630] bp at 90%. The latter includes both underperformance and a large positive contribution. It therefore supports further investigation of that setting, without establishing it as the most reliable choice.

The holdings also change substantially. At 20%, the factor replaces an average of 9.9 of the baseline’s 50 stocks; at 80%, it replaces 37.9. Model forecasts still rank stocks within each priority group, so these results measure their interaction with ECPND. Unscored holdings rise from 2.0 at 20% to 7.9 at 80% and 13.7 at 90%. All counts refer to entry holdings in the six-month sample. The next control examines whether positive means remain when those unscored stocks are unavailable to either portfolio.

Results also depend on the starting period. The 90% setting produces a six-month mean difference of -212 bp for 2024 formations and +916 bp for 2025 formations. A higher full-period average can therefore coexist with weak results in particular periods. These groups describe portfolio starting years, rather than calendar-year strategy returns.

10.3 Does the pattern depend on missing scores?

The first comparison leaves unscored stocks eligible without a penalty. As more scored stocks move into the lower-priority group, unscored candidates can take more portfolio places. The control tests this directly by removing missing scores before either portfolio is selected. Both arms then use the same scored-only universe. A numerical zero remains a valid score, and all other selection and execution rules are retained.

The same 232, 219 and 193 completed formations enter the three horizons, with four seeds each. Every holding has a score at entry. The final column counts lower-priority stocks used to fill the 50 places, averaged over the six-month sample. These are scored holdings, not missing values.

Positive mean differences remain at every tested threshold and horizon. At six months, the paired uplift is +63 bp at 20%, +118 bp at 50%, +293 bp at 80% and +478 bp at 90%. The positive pattern therefore also appears when every holding has a score at entry. Eligibility of unscored stocks is not required for positive mean uplift in this sample.

The reference portfolio changes too: its mean six-month return is 6.22% in the scored-only test, compared with 6.05% in the original universe. Each reported uplift belongs to its own matched comparison. Comparing the two uplifts does not isolate the return contribution of the excluded stocks.

The 90% setting needs an additional distinction. When fewer than 50 higher-priority stocks remain, the rule fills the available places from the lower-priority group: 3.6 holdings on average in the six-month sample. All have valid scores, but the portfolio is not restricted to the highest-scoring tenth. Its 95% interval, [-53, +1122] bp, also leaves substantial uncertainty around the larger mean.

Figure 4Selection strength and uncertainty

Left panel - stocks with and without scores. Both portfolios start from the full model-eligible universe. A stock can be selected even when no eligible ECPND value is available for the decision date. Unscored stocks are excluded from the score-percentile calculation and receive no factor penalty. As the threshold rises, more scored stocks move into the lower-priority group, allowing unscored candidates to occupy more places under the model ranking.

Right panel - scored stocks only. Stocks without an eligible ECPND value are removed before either portfolio is formed. The baseline selects the highest model-ranked stocks from this restricted set; the factor portfolio applies the ECPND priority rule to the same set. Every holding has a score at entry, with numerical zero still treated as a valid value. This tests whether positive mean uplift remains when no unscored stock can be held.

Reading the two lines. Each point shows the mean six-month factor-minus-baseline return at the indicated threshold. The left comparison uses the full-universe baseline; the right uses a scored-only baseline. The shared scale helps compare their results and uncertainty. The gap between the lines cannot be read as the return contribution of removing unscored stocks, because both the baseline and factor portfolios change.

10.4 What to test next

The next stage connects broader configuration tests with forward evaluation and research on a second earnings-call factor. The aim is to identify applications whose contribution persists across model choices, portfolio rules and newly observed periods.

Configuration and portfolio rules

The threshold results show that the way ECPND enters a portfolio matters. Further comparisons should use a predefined set of configurations, with the primary 20% rule retained as a reference. Three dimensions are particularly useful:

• Selection strength and coverage: Compare score thresholds in both the full and scored-only universes. Test lower-priority selection alongside strict exclusion, declaring how to handle fewer than 50 eligible stocks.
• Portfolio construction: Vary portfolio size, holding period and rebalancing frequency. Track return differences together with drawdown, turnover and costs, including continuous accounts that carry the same capital through time.
• Baseline and exposure controls: Repeat selected rules with independently chosen stock-ranking models. Match sector, size, momentum, volatility and liquidity exposures to assess what contribution remains.

Configuration choices should use only outcomes that have matured before the selection date, then remain fixed for the next evaluation period. All predefined settings should be retained in the record. Changing the score calculation itself requires a separate version, allowing application changes and measurement changes to be evaluated separately.

Forward live testing

Forward testing records a decision before its subsequent market outcome is known. The implemented ECPND protocol fixes the model and score version, uses the published end-of-day score snapshot and commits portfolio selections before the next eligible market opening. It retains the 20% rule, four model seeds and paired baseline/factor accounts, with both weekly cohorts and continuous accounts. Decisions, trades and valuations are stored as append-only records; missed decision dates are not filled retrospectively.

This is prospective paper evaluation. The internal ledger applies declared execution and cost rules; broker paper fills provide a separate execution record. Monitoring should cover snapshot availability, skipped decisions, return differences, drawdowns, turnover and execution costs. Three-, six- and twelve-month results enter the horizon comparisons as they mature. New thresholds, score versions or combined-factor rules should start separately identified forward tests, preserving the original ECPND reference and its history.

ECQDI as a complementary research input

ECQDI - Earnings Call Quantitative Dynamics & Intensity measures the quantitative character of external analysts’ questions relative to earlier calls in the same sector. It adds a text-based perspective to ECPND’s participation history. Their mean cross-sectional rank correlation is +0.076 in the combined study, indicating that they often rank stocks differently. This gives a reason to test complementary use; it does not establish statistical independence or improved portfolio returns.

The standalone ECQDI study finds positive mean buy & hold contributions, while continuous weekly results are weaker. Further work should therefore examine when its information is useful, how long to retain it and how frequently to act on it. Each factor’s own availability and expiry rules must be respected when joining the two inputs at a portfolio decision.

Combining participation and quantitative questioning

The completed combination study already tests 60 selection rules: different weights on the two score ranks, lower priority when either factor is weak, and lower priority only when both are weak. It also includes 52 ECPND-only controls that flag the same number of candidates as the corresponding combined rule. These controls help distinguish the second factor’s contribution from the effect of making selection more restrictive.

For example, a 75/25 combination gives 75% weight to ECPND’s current percentile rank and 25% to ECQDI’s rank. These are score weights, not capital allocations. At the 20% selection threshold, this rule produces +69 bp of six-month buy & hold uplift versus Qlib, compared with +54 bp for ECPND alone. Its incremental interval includes zero, its advantage weakens in later formations, and its continuous annualized growth remains below ECPND alone. The result identifies a configuration worth examining further, without establishing a dependable combined improvement.

The next combined evaluations should retain four references: the unchanged price-based model, ECPND alone, ECQDI alone and the combined rule. Coverage, costs and selection-strength controls should remain comparable. A small set of fixed combinations can then be evaluated prospectively, alongside independently chosen models, to test whether the second input adds value beyond the stronger single-factor alternative.

Research status. The completed threshold and combination experiments reuse previously studied historical data. Their intervals describe individual comparisons and do not adjust for selecting among configurations. The primary results in this note remain based on ECPND’s 20% rule. Forward results require their own start dates, protocol versions and maturity counts; the historical improvements reported here do not constitute a forward performance record.

AAppendix: Study evidence and verification

The evidence for this note consists of three sets of records: a recomputation of the portfolio study, a separate reconciliation of portfolio accounting, and two extensions of the selection-threshold tests. This appendix explains what each examination covers, how the reported results were checked and which evidence versions support this edition.

A.1 Frozen portfolio revalidation - 21 September 2026

The revalidation rebuilds the portfolio comparisons from the saved model forecasts, ECPND scores and adjusted opening and closing prices. These inputs are held fixed, so the exercise tests whether the stated portfolio rules reproduce the recorded results. The same evidence file supplies Figures 1-3 and Tables 2-5, including the primary comparisons, results by formation year, cost and eligibility checks, and continuous accounts.

All four previously stored mean return differences are reproduced within 10⁻¹¹ in return units: buy & hold and weekly rebalance, each at three and six months. This agreement checks the consistency of the calculations; it does not imply that future returns can be estimated to that precision. The revalidation also extends the study to twelve months, applying the same account rules and admitting only formations whose full holding period has elapsed.

The cohort run comprises 25,984 portfolio scenarios across the tested policies, seeds, costs and selection rules. A scenario is one simulated account path for a particular formation and configuration; that path can supply more than one completed horizon. These are repeated applications to the same market history, rather than independent market observations. The simulator records no incomplete scenarios, and the date-ordering check finds no model forecasts dated on or after execution. Transcript eligibility follows the separate timing convention in Section 2.

A.2 Internal ledger reconciliation - 20 September 2026

The ledger check approaches verification from the individual transactions. Its structured sample contains 80 portfolios, covering both portfolio policies, both comparison arms and all four seeds in each of the five formation years. Across that sample, 84,295 trades and 9,952 daily valuations are reconciled.

Starting with each account’s initial capital, the check rebuilds cash and share quantities from recorded purchases, sales and fees. It verifies trade notionals and trading costs, then values the resulting holdings at the frozen closing prices. Reconstructed net asset value is compared with the stored daily value. Horizon returns, maximum account values, drawdowns and their recorded dates are also checked against the daily series.

The largest absolute difference in daily net asset value is below $7 × 10⁻¹⁰, consistent with floating-point rounding. A separate completeness check confirms 16 portfolio records for every formation: four seeds, two policies and two arms. The transaction-level reconciliation covers the 80-account sample; it is not a claim that every stored account received that check. This is internal verification of accounting against the study inputs, not an external audit of market data, universe coverage or transcript publication records.

A.3 Threshold sensitivity and scored-only control - 22 September 2026

Two buy & hold sweeps vary the lower-priority score share from 10% to 90% in ten-percentage-point steps. The first retains stocks with missing scores under the primary rule. The second removes them before either portfolio is selected, creating a scored-only baseline and its matched factor portfolio. Forecasts, score definitions, prices, allocation and costs remain those of the primary study.

Each sweep simulates one baseline and nine factor variants for each of four seeds across the 232 formations with at least a completed three-month horizon: 9,280 account paths. Longer-horizon comparisons use only the subset that has matured. Neither sweep records incomplete scenarios, and every entry holding in the scored-only control has an eligible score.

Before extending the comparison, the first sweep reproduces the earlier 10%, 20% and 30% results at all three horizons. The control reproduces the earlier scored-only 20% result at those horizons. The checks compare mean baseline and factor returns, paired differences, interval endpoints and formation counts. Figure 4 and Tables 6-7 are generated from these extension records. They provide evidence about application choices within the studied period; the primary specification remains 20%, and the wider search is not a new out-of-sample test.

A.4 Published data and evidence versions

The research package separates observations from study outputs. The daily score CSV supports independent testing; the study-result and cohort files report our completed comparisons. A worked selection example shows how the model ranking and ECPND priority rule lead to candidate decisions. A short README explains the files and their use. Coverage and field definitions are described in Section 9.

The two-year score sample supports research over its supplied dates. Reproducing the full 2022-2026 study additionally requires the earlier score history, frozen forecasts, historical eligibility panel and price inputs used in the original run. The ECQDI companion folder contains separate evidence for that factor; the combination results discussed in the outlook are documented in the combined-factor note.

The numerical tables and charts are generated from the recorded result files. The SHA-256 checksums below identify the exact primary, reconciliation and exploratory evidence versions used here. A matching checksum confirms identical file contents, making it possible to distinguish this edition from a later recomputation. It identifies a record; the calculations and checks described above establish what that record contains.