Capstone Research Report: 18-Item Teaching Guide

After reading this guide, you should be able to connect the research question through monthly returns, signals, targets, drift, costs, paired differences, and inference. The README gives current status; the protocol retains the full mathematics; the time contract and monthly-return dictionary define fields and eligibility. This guide specifies future outputs. No target-market values have been computed; the design is fixed.

Source codes and reading references

The question, exposure, historical source-review scope, and metadata for S01/S07/S08/S09 are incorporated in Research Design and the read-only I3 source map. S02/S05 refer to the Protocol and French protocol config in the read-only I3 source; S03 is the Time Contract, and S04 is the Monthly-Return Dictionary. S06 is the accepted source mapping for this guide's 18-item structure. R/P refer to the original-main acceptance decisions and current pending status, available from the README and source map. These codes locate sources; readers do not need to consult QA outside the package.

1. Research Hypothesis

Question to understand: Turn a proposed mechanism into a falsifiable comparison. Prerequisites: Chapters 22–25; Projects P1–P3.

Fixed design and upstream source: S01/S02: fixed French ten industries, L11/skip1/top3, EW10; $H_0:\mu_d\le 0$ / $H_1:\mu_d>0$; one month is the experimental unit; slow information diffusion is not an identified cause.

Why and how: A fixed asset set, frequency, baseline, and directional hypothesis produce one d per month. The future primary table reports n, net mean, one-sided lower bound, two-sided interval, and the 10 bps economic threshold; fix the columns before viewing results. The mechanism is a hypothesis, with risk premia, concentration, and reversal as alternatives.

Future output and interpretation: User decision, RQ, and fixed primary endpoint → research question; the one-sided net mean difference answers the selected question. Target-market result fields: not run.

Failure branch: Changing the hypothesis or baseline after seeing results. Retain future results for all 132 months; do not replace them with a generic “hypothesis proposed” statement.

2. Data Sources

Question to understand: How one data identity limits conclusions. Prerequisites: Chapters 3 and 10; Projects P1–P2.

Fixed design and upstream source: S08/S09, R/P: 2026-10-05 revised snapshot, industry/FF3 identity, and download UTC; the actual subject/use for G01 is unknown, and raw redistribution is not_requested.

Why and how: First verify data/config/run identity and rights evidence for the actual activity; only then may inputs be read. Report snapshot, coverage, units, SHA, download UTC, and rights-status table. Full historical coverage does not mean the data were known at each historical point in time. Current materials contain metadata only; no new market files.

Future output and interpretation: Source text, license, and snapshot hash → authorized input; traceable source/use and revision scope. Target-market result fields: not run.

Failure branch: If licensing is unknown, do not execute real research; do not copy CSV/ZIP files. Source review remains within its historical scope; do not claim a new full-source review or rights grant.

3. Data Dictionary

Question to understand: Fields, keys, frequency, and units. Prerequisites: Chapters 2–3 and 11; Project P1.

Fixed design and upstream source: S04/S03: long-table key (snapshot_id, month, industry_id), decimal total returns for ten industries, FF3/RF, score/weight/amount/net-return/label-maturity fields.

Why and how: Illustrative research long-table schema: (snapshot_id, month, industry_id, total_return_decimal). This is a field example, not a market observation. Align FF3 by month; produce missing/duplicate/version logs. industry_id does not replace a permanent security ID.

Future output and interpretation: Fields, keys, and time units → valid monthly panel; consistent economic meaning. Target-market result fields: not run.

Failure branch: Duplicate keys or a second unit conversion. Future logs must record missingness and versions.

4. Point-in-Time Logic

Question to understand: A monthly label is not a historical publication clock. Prerequisites: Chapters 3, 12, 31, and 38.

Fixed design and upstream source: S03: half-open $[B_t,B_{t+1})$; $S=B_t-1\text{ second}$ and $\mathrm{allocation}=B_t$ are hypothetical ordering markers only; available / ingested / revision remain unknown. The normal and future rows are CONSTRUCTED.

Why and how: On the normal constructed row, first require the latest eligibility time $\le\mathrm{signal}$, then $\mathrm{signal}<\mathrm{trade}$; reject the future row when eligibility is later than signal. Separately test label_end / available / ingested against fit cutoff $F$, then test half-open interval overlap. Adjacent labels can still be immature even if they do not overlap.

Future output and interpretation: Eligibility table and normal/future counterexample → input-eligibility status. The constructed PIT rule is testable; historical PIT remains unknown. Target-market result fields: not run.

Failure branch: Calling revised history PIT-valid or admitting a future label into training. Do not fill market history with skip1, download UTC, or constructed second-level timestamps.

5. Data Cleaning

Question to understand: What is rejected and why no silent filling is allowed. Prerequisites: Chapters 3 and 29–31.

Fixed design and upstream source: S02/S04/S06: reject duplicate month/industry keys, missing months, non-finite values, and missing values; convert original percentages to decimal by dividing by 100 once only; do not add corporate-action cash already included in total return.

Why and how: For every month, verify ten unique industry keys, a complete continuous month sequence, finite decimals, missing-value sentinels, and the one-time percentage conversion. Output rejection reasons and the retained schema; do not zero-fill, replace industries, or silently drop rows. Corporate-action cash is not added again to total returns.

Future output and interpretation: Schema counts and missing-sentinel log → rejected/retained lists; processing remains traceable. Target-market result fields: not run.

Failure branch: Silently dropping months, zero-filling, or applying the 3,600-second price-staleness rule to monthly returns. Do not fabricate future count tables.

6. Feature Engineering

Question to understand: Compounding endpoints and units. Prerequisites: Chapters 4, 11, and 25.

Fixed design and upstream source: S02/S03/S05: $\operatorname{score}_{i,t}=\prod_{m=t-12}^{t-2}(1+R_{i,m})-1$, 11 terms; general window $[t-\mathrm{skip}-L,\ t-\mathrm{skip}-1]$; all first-evaluation-month windows are listed.

Why and how: Each (1+R) is a wealth multiplier; multiply 11 terms and subtract 1. Summing returns omits cross-products. For the first evaluation month, 1990-01, L11 uses 1989-01..1989-11 and skips 1989-12. A future score table records t, input_start/end, all 11 terms, score, and input identity.

Future output and interpretation: 11-month compounding formula, endpoints, and constructed expansion → score, using exactly $t-12,\ldots,t-2$. Target-market result fields: not run.

Failure branch: Treating a sum as compounding or using the holding-month return in the signal. Use the existing constructed example; do not recompute French data. The future output is a ten-industry score table.

7. Signal Construction

Question to understand: How deterministic sorting produces top-k. Prerequisites: Chapters 23, 25, and 28.

Fixed design and upstream source: S02/S05: fixed ten-industry order; descending score and ascending manifest index for ties; primary top3 at $1/3$ each; each sensitivity assigns $1/k$; no fitting.

Why and how: Sort the ten scores uniquely by (score descending, manifest index ascending). Assign $1/k$ where rank $\le k$, otherwise 0. Output all ten ranks/targets, not only winners. Do not winsorize, standardize, or neutralize; any added fit requires a protocol revision.

Future output and interpretation: All ten industry scores and fixed tie order → top3; ranking is deterministic and not trained. Target-market result fields: not run.

Failure branch: Choosing industries after viewing future results or using hidden randomness for ties. The future rank/target table remains pending authorization.

8. Portfolio Construction

Question to understand: Targets, drift, and post-cost weights. Prerequisites: Chapters 9, 20–21, and 29; Project P3.

Fixed design and upstream source: S02: $V_{post}=V_{pre}-c\sum_i|w_iV_{post}-x_i|$; $x_i$ is the drifted holding amount; fully invested long-only with gross=net=1; first-month value $1/(1+c)$; no segment reset.

Why and how: Carry forward $x$ and $V_{pre}$, solve $V_{post}$, then trade $wV_{post}-x$; fees and NAV use the same amount unit. Check $\sum_iw_i=1$, long-only exposure, cash conservation, and month-end drift. Future ledger columns: pre/post/end NAV, x, w, trade, fee, terminal. The full equation is in the protocol.

Future output and interpretation: Targets/drift table and self-financing equation → ledger; the long-only budget closes each month. Target-market result fields: not run.

Failure branch: Mixing pre/post-fee weights or drift; computing costs from something other than the sum of absolute buy/sell amounts; missing the target weights after costs; omitting entry/terminal costs; or resetting at segment boundaries all break cash conservation. The original engineering share example is a unit comparison only, not a French adapter.

9. Transaction Cost Model

Question to understand: Trading-notional costs and NAV drag. Prerequisites: Chapters 29 and 34–36; Project P4.

Fixed design and upstream source: S02/S05/S06: all-in cost is $c\times|\text{traded notional}|$, where $c=0.001$; do not divide by 2 or charge spread again. $200.10\times0.001=0.20010$; $(1-0.001)(1+0.02)-1=0.01898$ is a separate constructed amount example.

Why and how: $c=10/10000=0.001$; USD 200.10 traded notional produces USD 0.20010 all-in cost. NAV-cost proportion $C=0.001$ and gross return $r=0.02$ produce net $(1-C)(1+r)-1=0.01898$, a drag of $0.00102=10.2$ bps; C and c have different denominators. The future cost bridge uses month on the x-axis and one consistent amount unit or an explicitly stated NAV proportion on the y-axis; do not mix the two.

Future output and interpretation: Absolute traded notional / fee / gross-to-net bridge → net ledger; reconcile NAV under the cost scenario. Target-market result fields: not run.

Failure branch: Dividing by 2 and understating cost, charging spread twice, or claiming real capacity without data. The share example's half-spread/commission is another setup. Future cost bridge and capacity remain unknown.

10. Walk-Forward Validation

Question to understand: Fixed ranking still requires an access and phase contract. Prerequisites: Chapters 22, 31, and 38; Projects P3/P6 are comparisons only.

Fixed design and upstream source: S02/S03/S05/S07: 18 warm-up months; $440=176+132+132$; sole primary segment 2015-09..2026-08; 13 continuous paths carry across boundaries; do not invent training; the history is exposed and is not a blind holdout.

Why and how: Assign rows using the TIME_CONTRACT phases: 176 development, 132 validation, and 132 locked-reporting months, total 440. This separates display/reporting only. The first primary row carries from 2015-08; the last primary row includes liquidation. Output a month/phase/access/hash table. A fixed ranking has no training; skip1 is not embargo.

Future output and interpretation: Confirmed 176/132/132 months, continuous ledger without reset, and maturity/purge table → 132-row locked report. Fix access and order first; ranking does not require invented training. Target-market result fields: not run.

Failure branch: Calling exposed history a holdout or running with unresolved window rules. Run/code/config/data/hash/access logs require authorization.

11. Performance Analysis

Question to understand: Difference between the primary net-difference test and descriptive statistics. Prerequisites: Chapters 7, 11, and 32.

Fixed design and upstream source: S02/S05: $d_t=r_{M0,net,t}-r_{EW10,net,t}$; primary $n=132$, $k=1$, common-$n$ denominator for $\widehat\gamma_j$, Bartlett lag 6, $q=132/131$ applied to covariance, $\mathrm{SE}=\sqrt{\widehat\Omega/131}$, approximate $t(131)$; bootstrap block 12 / 5,000 / fixed seed 20261012.

Why and how: Check continuous month keys for the primary 132-row d, then compute HAC6 and the fixed bootstrap. The primary interval plot uses net mean difference on the x-axis (decimal/month or clearly labeled bps/month) and specification on the y-axis, with 132 months, hash, and one-/two-sided intervals identified. The wealth plot uses month on the x-axis and NAV on the y-axis, for all 440 months descriptively only. CAGR differs from 12 times the monthly mean. If SE degenerates, do not report a p-value.

Future output and interpretation: Paired monthly difference/HAC6 for the 132 locked-reporting rows; all 440 months and the first two phases are descriptive only → primary and secondary results. Separate net advantage from uncertainty. Target-market result fields: not run.

Failure branch: Showing only the highest Sharpe or best interval. Section 11 explains one-/two-sided intervals, degenerate SE, no fabricated p=0 for missing months, and the 10 bps economic threshold. Provide future table columns and interpretation rules only; do not generate means, Sharpe ratios, or CIs now.

12. Risk Analysis

Question to understand: Concentration, tails, and missing volume beyond returns. Prerequisites: Chapters 18, 20–21, and 33–35.

Fixed design and upstream source: S01/S02/S06: top3 concentration, industry weights, maximum drawdown, and historical tails. A uniform constructed $-10\%$ shock checks only the risk formula; with no volume/order book, do not estimate ADV capacity.

Why and how: Interpret concentration in weights, maximum drawdown, and historical tails separately. Weight chart: month on x-axis, capital weight on y-axis. Drawdown chart: month on x-axis, loss relative to peak on y-axis. Label the monthly frequency/window for tail measures. The uniform constructed $-10\%$ shock checks only the risk formula; it does not measure market liquidity or ADV capacity.

Future output and interpretation: Industry weights/drawdown/tails/constructed shock → risk figures; explain return trade-offs, concentration, and loss. Target-market result fields: not run.

Failure branch: Estimating capacity without volume or claiming equal weights eliminate concentration risk. Mark future weights/drawdown/tails as not run; do not use the engineering USD 200 stress as a market-risk estimate.

13. Factor Attribution

Question to understand: Post-realization attribution and one RF subtraction. Prerequisites: Chapters 14–15; Project P2.

Fixed design and upstream source: S02/S05: subtract same-month FF3 RF once from net total return; do not subtract RF again from published Mkt-RF / SMB / HML. Include an intercept and use $k=4$ correction $n/(n-4)$. The primary paired $d_t$ does not subtract RF and uses $k=1$.

Why and how: Set $y_t=\texttt{net\_total}_t-\texttt{FF3\_RF}_t$ once, and let $X$ contain an intercept plus published Mkt-RF / SMB / HML; use $k=4$ correction. Future tables report monthly alpha, beta, and HAC intervals. Residual chart: month on x-axis, decimal/month on y-axis. The primary $d_t$ does not subtract RF; alpha is post-hoc attribution conditional on the specified model.

Future output and interpretation: Same-month net total return minus RF once / FF3 k=4; primary difference has no RF subtraction and k=1 → coefficient/residual figure; historical model-related attribution. Target-market result fields: not run.

Failure branch: Mixing FF3/FF5, subtracting RF twice, or treating alpha as skill. Future coefficients/residual figure remains pending; do not mix FF3 and FF5 snapshots.

14. Robustness Tests

Question to understand: How to report all prespecified variations. Prerequisites: Chapters 22 and 30–32; Project P3.

Fixed design and upstream source: S02/S05/S06: report cost, skip, fixed-phase, and statistical block-length 6/24 sensitivities in full. Keep negative and zero results; do not choose conventions based on intervals.

Why and how: Report every prespecified cost/skip sensitivity, the fixed three calendar phases, and bootstrap 6/24-month sensitivities. Future table columns include specification, fixed variation, window, mean, interval, and status. Segments are cuts through a continuous ledger. Report conflicting intervals and negative/zero results; do not select a favorable convention.

Future output and interpretation: Prespecified costs/skips/three prespecified phases → complete table; describe changes under fixed variations. Target-market result fields: not run.

Failure branch: Dropping negative/zero scenarios or adding and reporting only favorable post-hoc items. Fill in prespecified table columns and interpretation rules now; leave actual results blank.

15. Parameter Sensitivity

Question to understand: Why 9+4 is not a search contest. Prerequisites: Chapters 9, 22, 25, and 32.

Fixed design and upstream source: S02/S05: nine strategies M0, C0, C25, C50, L5, L17, K2, K4, and S2 plus four EW cost paths; zero optimization. Vary one factor at a time; do not make a Cartesian grid.

Why and how: List all 13 paths explicitly. A one-factor response plot uses cost in bps, L, skip, or k on the x-axis and descriptive net difference or risk on the y-axis, labeled with input identity and continuous window. There is no Cartesian product and zero optimization; do not replace M0 with the best point.

Future output and interpretation: Trial ledger for 9 strategies + 4 benchmarks → one-factor plots; describe response stability without selecting a winner. Target-market result fields: not run.

Failure branch: Exceeding the budget with a Cartesian product or replacing primary parameters after seeing results. Future response plots do not select the primary specification.

16. Overfitting Discussion

Question to understand: Fixed future procedures do not erase exposure. Prerequisites: Chapters 22, 30, 32, and 37; Projects P1–P3/P6.

Fixed design and upstream source: S07/S02: P3 saw the full 1990–2026 period; P1/P2 saw portions; other access is unknown. Not blind; full_preregistration=false. PSR/DSR/PBO may be N/A if applicability is insufficient.

Why and how: Record exposure to P3's full-history equal-weight/model results and partial exposure in P1/P2; the 9+4 degrees of freedom are fixed. Label future exploration as post-hoc and preserve the complete failure log. If evidence for applying PSR/DSR/PBO is insufficient, report N/A or not_evaluated rather than fabricate a statistic.

Future output and interpretation: Exposure/degrees of freedom/total attempts → interpretation boundary; fixed steps cannot remove hindsight bias from prior exposure. Target-market result fields: not run.

Failure branch: Filling in DSR/PBO without applicable evidence. Mark future exploration as post-hoc; do not invent DSR/PBO results.

17. Limitations

Question to understand: What this route cannot directly answer. Prerequisites: Chapters 3, 29, 33–36, and 45.

Fixed design and upstream source: R/P, S02/S04/S10: rights are unresolved, historical PIT is unknown, hypothetical allocations are not real fills or ETF replication, and costs are not calibrated to capacity. Retain old P6/39 gaps; the monthly-return adapter is not implemented.

Why and how: The support-domain table can be filled now: revised history, monthly allocation, and assumed linear costs can be designed; strict historical PIT, ETF replication, real fills, capacity, and causal mechanism cannot be established directly. G01 is unknown, G05 is not authorized, the monthly-return adapter is not implemented, and real-data gaps from P6/39 remain.

Future output and interpretation: Rights/PIT/trading/statistical gaps → support-domain table; define what the study can and cannot establish. Target-market result fields: not run.

Failure branch: Turning unknown into zero or adding corporate-action cash on top of total return. Unknown stays unknown; engineering acceptance does not mean real research is complete.

18. Final Research Conclusion

Question to understand: Separate the design answer from the future research answer. Prerequisites: Chapters 22, 32, and 45; all prerequisite material.

Fixed design and upstream source: R/P, S01/S02/S06: reportable current state is design/protocol technically accepted and execution blocked. Future positive, zero, negative, or interval-crossing-zero outcomes have distinct interpretations. An engineering PASS does not answer $\mu_d>0$.

Why and how: Current answer: the research design and technical protocol are accepted, but there are no target-market results, so it is not possible to judge whether $\mu_d>0$. A future positive result with lower bound $>0$ must still be compared with the 10 bps economic threshold; an interval crossing zero means insufficient evidence; net $\le0$ means no net advantage for this specification; gross advantage consumed by costs indicates cost sensitivity; a failed gate means blocked, not zero effect.

Future output and interpretation: Primary paired net results / failed gates → answer the RQ. The result may be negative, zero, or insufficiently precise; blocked means no market conclusion. Target-market result fields: not run.

Failure branch: Treating engineering success as the result of all 18 research items. Provide future conclusion wording and failure branches, and never use constructed tables or engineering NAV 9995.998 as a French research conclusion.