Capstone 3 — Innovation from USPTO PatentsView

ASSIP framing — read first. The ASSIP 2026 terminal deliverable is a research poster (~48”×36”) + a JSSR abstract (200–300 words) + an 8-minute poster talk + a replication packet — not a full paper. This exemplar is the full research narrative the poster summarizes: a complete empirical-finance paper written one level above the ASSIP graded artifact, so that you can see what the full argument looks like before you compress it onto six poster panels and a 10-sentence abstract. Use it as a structural source for your ASSIP poster (each section maps to a panel) and, if you choose to write the optional companion full paper, as a length-and-shape template for that paper.

Illustrative results — read this first. Every number in this paper comes from synthetic patent and stock-return data constructed for instruction in the Week-6 companion notebook nb6.1 (notebooks/week-06/nb6.1-patent-value-panel.ipynb). The data are simulated from a seeded random generator with a known “true” patent value baked in, precisely so the construction can be checked against a ground truth no real dataset gives you. These are not empirical findings. No claim here describes the actual U.S. patent system, any real firm, or any real return. The paper is a gold-standard template: it shows what a finished innovation-measurement capstone looks like when it is built to the camp’s standards (CONVENTIONS §4, Appendix D), with the real-data pull documented but not executed. Treat the magnitudes as worked-example outputs, not evidence.

Reading a Patent’s Worth Off the Stock Price: A KPSS-Style Innovation Measure and a Grant-Date Event Study

Leah R., ASSIP 2026 (Aspiring Scientists Summer Internship Program)

Abstract

How much is a single invention worth? Patent counts treat a trivial design tweak and the transistor as one unit each; forward citations are slower and arrive a decade too late to value a recent patent. Following Kogan, Papanikolaou, Seru & Stoffman (2017), I build a market-based, dollar-denominated innovation measure: the abnormal stock return in a three-day window around a patent’s grant date, scaled by market capitalization, read as the market’s real-time estimate of the patent’s private value. Working in a seeded synthetic panel of 60 firms and 633 patents over 2018–2021 — where the true patent value is known by construction — I reconstruct the measure with a Week-4 market-model event study and then run two tests. First, a recovery check: the reconstructed value correlates with the planted true value at 0.57 in levels and 0.38 in logs at the patent level, rising to 0.72 once aggregated to the firm-year, because averaging cancels the idiosyncratic window noise. Second, a predictive check: regressing next-year firm profitability on this year’s log innovation value yields a positive, significant coefficient (\hat\beta = 0.0054, t = 3.6), the same validation pattern KPSS report on real data. A contamination experiment — injecting simulated earnings surprises into a rising share of grant windows — degrades both the recovery correlation (0.38 to 0.25) and the predictive t-statistic (3.6 to 2.5), making visible the confounding-news threat that the short window and firm-year aggregation are designed to blunt. The exercise demonstrates how to turn an unobservable (the economic value of an idea) into an observable, and exactly where the construction can break.

1. Introduction

The single most cited patent of the 2010s and a patent that protects a slightly rounder bottle cap both count as “one patent.” That is the problem that has dogged the economics of innovation since it began: the obvious way to measure how much a firm invents — count its patents — weights a breakthrough and a triviality equally, and the standard fix — count the forward citations a patent later attracts — is both slow, since citations accrue over ten or more years, and truncated, since a patent granted last year looks uncited only because the future has not happened yet. Neither proxy is denominated in the unit anyone making a capital-allocation decision actually cares about, which is dollars. An R&D manager, an investor, or a policymaker wants to know what an invention is worth, not how many of them there are.

This paper builds and stress-tests a dollar-denominated innovation measure, in the spirit of Kogan, Papanikolaou, Seru & Stoffman (2017, hereafter KPSS), on a synthetic dataset where I know the right answer. The contribution is pedagogical and methodological: I demonstrate, against a known ground truth, that a patent’s private economic value can be read off the stock-market reaction to its grant, that the resulting firm-year measure predicts future profitability, and that confounding news in the grant window degrades the measure in a direction and magnitude I can quantify. Because the data are simulated with a planted true value, I can do what no empirical paper can — compare the constructed measure to the truth it is trying to capture — which is the whole point of building it in a sandbox before trusting it in the wild.

I find three things. First, the event-study reconstruction recovers the planted patent value: the correlation between estimated and true value is 0.57 in levels and 0.38 in logs across patents, and climbs to 0.72 when patents are aggregated to a firm-year panel, because firm-year sums average out the day-to-day idiosyncratic noise that contaminates any single three-day window. Second, the measure validates the way KPSS’s does: regressing next-year firm profitability on this year’s log innovation value gives a positive and statistically significant coefficient (t = 3.6), tracking the same predictive relationship the published measure exhibits on real firms. Third, the measure breaks in a legible way: when I contaminate a growing fraction of grant windows with simulated earnings-announcement returns, both the recovery correlation and the predictive t-statistic fall monotonically, exposing the confounding-window threat that is the central vulnerability of the design.

The paper proceeds as follows. Section 2 positions the measure against the patent-counting and citation literature and states what KPSS added. Section 3 describes the synthetic data and the real-data pull it stands in for. Section 4 lays out the event-study design, states the identifying assumption in one sentence, and tabulates the threats to it. Section 5 reports the headline recovery and validation results. Section 6 runs the robustness checks — window length, a placebo window, and the contamination sweep. Section 7 concludes with the limits I cannot design away.

2. Positioning

The literature on measuring innovation is a history of progressively better proxies, each fixing a defect in the last. The first generation counts patents: available far back in time, but treating every patent as equal, so a firm with a hundred trivial patents outscores the firm holding the one patent that defined a decade. The second generation, the workhorse of the field, counts forward citations — how many later patents cite this one — which genuinely tracks importance (the lineage runs through the NBER Patent Data Project of Hall, Jaffe & Trajtenberg). But citations are slow, accruing over a decade so a recent patent cannot be valued in real time, and truncated, since recent cohorts mechanically show fewer citations simply because less future has elapsed — making cross-cohort comparison the classic apples-to-oranges error.

KPSS’s move, which this paper imitates, is to let a third party that has already done the valuation work — the stock market — supply the number. When the U.S. Patent and Trademark Office grants a patent to a publicly traded firm, the grant is news: it resolves uncertainty about whether and on what claims the patent would issue. If the patent is valuable, the firm’s stock should jump when the grant becomes public; if it is worthless, the stock should not move. KPSS therefore read the abnormal return in a short window around the grant date as the market’s dollar-denominated estimate of the patent’s private value, and released the resulting patent-level estimates publicly, which is why the measure has become standard infrastructure across finance, economics, and management. This is the same event-study logic Week 4 applied to corporate events; here the “event” is a patent grant, and the abnormal return is the valuation. My contribution relative to KPSS is not a new measure but a transparent reconstruction: by building it on data with a known answer, I show exactly how much signal the method recovers and how much it loses — which the real measure can never reveal, because the truth is unobserved.

3. Data

The analysis runs on a seeded synthetic panel generated in nb6.1 (random generator np.random.default_rng(42)), designed to mirror the schema of the real pull it stands in for. I simulate 60 firms over four calendar years, 2018–2021, at a daily frequency of roughly 252 trading days per year. Each firm has a market capitalization drawn on a log scale (median near $2,000m, heavily right-skewed), a market beta near one, and an idiosyncratic daily volatility between 1.0% and 2.0%. Daily returns follow a market model, R_{it} = a_i + b_i R_{mt} + \varepsilon_{it}, the same normal-return model the event study will invert.

Onto this return panel I graft 633 patents, roughly two to three per firm-year, each carrying a grant date and a true dollar value drawn from a right-skewed log-normal distribution whose scale rises with a latent firm-quality term. The resulting value distribution is wildly skewed in exactly the way real patent values are: the top 10% of patents hold 59% of the total planted value — the blockbuster phenomenon that is the entire motivation for not counting patents equally. Each patent’s true value is then injected into the data as a small abnormal return spread over its three-day grant window, scaled so that a fixed dollar value is a large percentage move for a small firm and a tiny one for a giant (a $50m patent moves a $200m firm by 25% but a $20,000m firm by 0.25%). This is the mechanism the event study must run in reverse.

The synthetic panel deliberately reproduces the selection structure of the real data. Only patents whose firm has enough pre-grant return history to estimate a market model receive a value; the rest drop out, exactly as private firms, foreign assignees, and individual inventors drop out of the real measure for lack of a CRSP listing. After this screen, the patent-level estimates aggregate to a 217-row firm-year innovation-value panel, the central data object.

In a real build, the data would come from three stitched-together sources, documented in the data cards and in a non-executed (RUN_REAL=False) cell of nb6.1. Patents, grant dates, and citations would come from USPTO PatentsView via the USPTO Open Data Portal (api.uspto.gov, X-API-KEY header; PatentsView migrated to ODP on March 20, 2026), or the bulk PatentsView tables on ODP for panel-scale work. The patent-to-firm link and the published per-patent values would come from the NBER-lineage crosswalk and the public KPSS patent-value file (bulk downloads, no key). Daily firm returns would come from CRSP, which is licensed and must stay read-only on GMU infrastructure with a pinned snapshot date (CONVENTIONS §5). The join chain is PatentsView patent → crosswalk → permno/gvkey → CRSP/Compustat → firm outcomes. I name this pipeline so the template is reproducible, but I do not execute it, and no real number appears anywhere in this paper.

4. Empirical design

4.1 Constructing the measure

The measure is built one patent at a time with a market-model event study, the Week-4 machine pointed at a patent grant. For patent j owned by firm i and granted on trading day t_0:

Estimate the normal-return model. On a pre-event estimation window of 120 trading days, ending 10 trading days before the grant (the gap prevents the event from leaking into the parameters), fit R_{it} = \hat a_i + \hat b_i R_{mt} by least squares.
Compute the abnormal return. On each day t of the three-day grant window (days t_0 through t_0+2), the abnormal return is AR_{it} = R_{it} - (\hat a_i + \hat b_i R_{mt}), the part of the return the market model did not predict.
Cumulate and convert to dollars. Sum the window abnormal returns into a cumulative abnormal return, \mathrm{CAR}_{ij} = \sum_{t=t_0}^{t_0+2} AR_{it}, and multiply by the firm’s market capitalization to turn a percentage into a dollar value: \widehat{V}_{ij} = \mathrm{CAR}_{ij}\times \text{MktCap}_i.
Aggregate. Sum the (positive) per-patent dollar values within each firm and year to form the firm-year innovation-value panel.

The one-line specification, in the camp’s discipline (CONVENTIONS §4): outcome (constructed) = dollar value per patent; key input = idiosyncratic stock return in the three-day grant window; adjustments = market-model-filtered, scaled by market cap; sample = patents granted to firms with usable pre-grant return history; the identifying assumption, in one sentence = the abnormal stock return in the three-day grant window equals the market’s unbiased estimate of the patent’s economic value, because markets impound the patent’s value at grant and no other value-relevant news contaminates the window.

For the predictive validation, the specification is a firm-year regression: outcome = next-year firm profitability; key regressor = log of this year’s innovation value; controls = log market capitalization; fixed effects = year; standard errors = classical (the synthetic panel is independent across firms by construction, so there is no within-firm clustering to correct; on real data this regression would cluster by firm, per CONVENTIONS §3); sample = the 217-row firm-year panel. This relationship is predictive, not causal — I do not claim innovation causes profits, only that the measure forecasts them the way a valid value measure should.

4.2 Threats to identification

The identifying assumption has three soft joints, each inherited directly from KPSS’s “what’s vulnerable.” Table 1 names each threat, its direction, and the design response.

Table 1 — Threats to the grant-window value measure, and responses

Threat	Why it biases the measure	Direction	Design response
Confounding window news (earnings, deals, analyst revisions landing in the 3-day window)	The CAR mixes the patent’s value with unrelated news, so \widehat{V} no longer isolates the patent	Either sign per patent; attenuates the measure on average if idiosyncratic	Short (3-day) window limits exposure; firm-year aggregation averages out idiosyncratic confounds; quantified in §6.3
Selection to granted patents of public firms	Only granted patents of firms with return history get a value; rejected/abandoned patents and private, foreign, university assignees are invisible	Bounds the population, not the within-sample estimate	Disclosed, not corrected; restricts external validity (stated in §7)
Expectations, not realized value	The window return is the market’s forecast at grant, not what the patent ultimately earned; bias if the market is systematically optimistic about flashy patents	Inflates value of patents the market over-rates	None available within the design; the measure is “market’s expected value at grant,” and prose says so
Efficient-markets assumption baked in	The measure assumes prices impound value promptly; if markets are slow the window misses leaked value	Attenuates	Short window plus sensitivity to window length (§6.1); cannot test efficiency with a measure that assumes it

The first threat is the one I can manipulate in a synthetic world, so it is where §6.3 aims its experiment. The remaining three I disclose and bound rather than solve, which is the honest posture for a measurement design.

5. Results

The headline result is the validation table: does the measure, reconstructed from three-day stock blips, both recover the planted truth and predict future outcomes? Table 2 reports both checks, with the true-value regression as a benchmark ceiling.

Table 2 — Recovery of the planted value and prediction of next-year profitability

	Recovery: corr(true, estimated)		Prediction: next-year profitability on log innovation value
	Level	Log	\hat\beta	SE	t
Estimated measure (event-study CAR × mktcap)	0.565	0.382	0.0054	0.0015	3.61
Firm-year aggregate (estimated)	0.722	—
Benchmark: true planted value	1.000	1.000	0.0152	0.0017	9.20

Notes. Synthetic data from nb6.1; illustrative, not empirical. 633 patents and a 217-row firm-year panel, 60 firms × 2018–2021. “Recovery” columns report the Pearson correlation between the event-study-estimated patent value and the known planted true value, in levels and in logs (logs computed only where the estimated value is positive, since a CAR can come out negative). The firm-year row aggregates patent values within firm-year before correlating. The prediction columns report a firm-year OLS of next-year profitability on log innovation value, controlling for log market cap and year fixed effects; the coefficient on log innovation value is shown. Standard errors are classical (firms are independent by construction in the synthetic panel; a real-data version would cluster by firm). The benchmark row substitutes the true planted value for the estimated one, giving the ceiling the estimate would hit with zero measurement error. ^{\*\*\*}\,p<0.01. Coefficient 0.0054 has t=3.61, p=0.0004.

Three readings. First, the event study recovers the truth, imperfectly but unmistakably. The patent-level correlation of 0.565 in levels (0.382 in logs) is far from perfect, and it should be: the median patent’s true value is a window blip of well under 1%, while a firm’s three-day idiosyncratic noise runs several percent, so any single window is a noisy read. This is not a flaw to apologize for; it is the realistic signal-to-noise ratio that motivates the next number.

Second, aggregation is the cure for window noise. Summing patents to the firm-year lifts the recovery correlation to 0.722, because the idiosyncratic noise in individual windows is mean-zero and cancels when many patents are pooled, while the signal accumulates. This is the mechanical reason KPSS-style measures are most trustworthy at the firm-year level — the level at which the economics is actually done — and least trustworthy as a single patent’s headline dollar figure.

Third, the measure predicts the future the way a valid value measure should. The estimated log innovation value carries a positive, significant coefficient on next-year profitability (\hat\beta = 0.0054, t = 3.61). The benchmark row shows the ceiling: with the true value substituted in, the coefficient is larger (0.0152) and the t-statistic far higher (9.20). The gap between the two rows is precisely the cost of measurement error — the estimated measure attenuates the relationship toward zero, exactly as classical measurement error in a regressor predicts, but does not destroy it. A three-day stock reaction, filtered and scaled, forecasts a real outcome a year out.

6. Robustness

A measurement design earns trust by surviving perturbations of the choices that built it. Three checks: the window length, a placebo window, and the confounding-news experiment that is the point of working in a sandbox.

6.1 Window length

The three-day window is a choice, and the right way to disclose a choice is to vary it. Shortening to a one-day window (0,0) throws away abnormal return that lands on days t_0+1 and t_0+2, attenuating the measure; lengthening to a five-day window (0,4) admits more unrelated news, raising the variance of the estimate and inviting the confounding-window threat of §6.3. The three-day window is the standard compromise — long enough to capture a reaction that does not all land on the grant date, short enough to keep the confound window narrow. The qualitative lesson, which nb6.1’s Your Turn extension makes a student reproduce, is that the recovery correlation is not invariant to the window; reporting the measure without disclosing the window would hide a real degree of freedom.

6.2 Placebo window

The cleanest falsification is a placebo: run the identical event study on a window where no patent value was injected — a stretch of trading days unconnected to any grant. By construction the synthetic data plant value only in true grant windows, so a placebo window should return a CAR that is mean-zero noise and an estimated value indistinguishable from zero, with no correlation to any patent’s planted value. A placebo that instead produced apparent “value” would signal that the event-study machine was manufacturing signal out of ordinary volatility — the single most important thing to rule out before trusting the real-window estimates. The design passes this check by construction here; on real data it is the test that separates a genuine grant-window reaction from a mechanical artifact, and it is the one I would run first.

6.3 Contaminating the grant windows

The headline robustness exercise attacks the design at its weakest joint (Table 1, row 1): confounding news in the window. I inject a simulated one-day earnings surprise — a normal shock with a 6% standard deviation, large relative to the patent’s own window signal — into a rising fraction of grant windows, rebuild the measure on the contaminated data, and watch both the recovery correlation and the predictive t-statistic degrade. Results are in Table 3, averaged over five contamination seeds for smoothness.

Table 3 — Confounding window news degrades both recovery and prediction

Share of windows contaminated	0%	15%	30%	50%	75%
Recovery correlation (log)	0.382	0.353	0.330	0.291	0.253
Validation t (next-year profit)	3.61	3.40	3.13	2.81	2.45

Notes. Synthetic data from nb6.1; illustrative, not empirical. A one-day earnings surprise (Gaussian, SD = 6%) is added to the first day of a random fraction of grant windows; the value measure is then rebuilt from the contaminated returns. “Recovery correlation (log)” is corr(true, estimated) at the patent level; “Validation t” is the t-statistic on log innovation value in the next-year-profitability regression of Table 2. Each column averages five contamination seeds. The 0% column reproduces the clean estimates of Table 2.

The pattern is monotone and interpretable. As contamination rises from 0% to 75%, the recovery correlation falls from 0.382 to 0.253 and the predictive t-statistic falls from 3.61 to 2.45 — the measure is being eaten by noise the event study cannot distinguish from the patent’s own signal. Two lessons follow. First, this is why the short window and firm-year aggregation are the defenses: a narrow window gives confounds less room to land, and aggregation averages out idiosyncratic confounds like a one-off earnings surprise. Second, and this is the subtler point KPSS’s “what’s vulnerable” insists on, aggregation only saves the measure when the confound is idiosyncratic. The dangerous case — flagged in nb6.1’s extension and not run here — is systematic contamination, where good-news days correlate with grant dates; that bias does not wash out under aggregation and would survive into the firm-year panel.

7. Conclusion

This paper reconstructs a KPSS-style, dollar-denominated innovation measure on synthetic data with a known answer, and uses that answer to show three things the real measure can never directly show: the event study recovers the planted patent value (correlation 0.57 at the patent level, 0.72 once aggregated to the firm-year), the measure predicts next-year profitability (t = 3.6, against a true-value ceiling of t = 9.2), and confounding window news degrades both recovery and prediction in a monotone, quantifiable way. The exercise turns an unobservable — the economic value of an idea — into an observable, and it locates the exact seams where the construction can fail.

The limits are real and I state them plainly. The measure is the market’s expected value at grant, not realized value; a patent the market loved at grant and that later flopped still scores high, and any sentence dropping the words “expected” and “the market’s” overclaims. The measure assumes efficient markets, so it cannot be used to test efficiency without circularity, and it is sensitive to the choice of normal-return model and window length (§6.1). The sample is selected to granted patents of firms with usable return history; the real-data analogue excludes private firms, foreign assignees, universities, and de-listed firms, so the firm-year panel describes the innovation of large, listed, surviving firms and nothing I show here generalizes to the whole economy. And the predictive validation is an association, not a causal effect — innovating firms grow and the measure forecasts it, but I have no random assignment of patents, no instrument, no discontinuity, and reverse causality (growing firms patent more) and common drivers (a demand shock lifting both) remain live alternatives. Above all, the data are synthetic. The numbers demonstrate that the method recovers a known truth and breaks under a known confound; they say nothing about real patents, real firms, or real returns. That separation — between validating a method and discovering a fact — is the discipline the whole exercise is meant to teach.

References

Kogan, L., Papanikolaou, D., Seru, A., & Stoffman, N. (2017). Technological Innovation, Resource Allocation, and Growth. Quarterly Journal of Economics, 132(2), 665–712.
Hall, B. H., Jaffe, A. B., & Trajtenberg, M. (2001). The NBER Patent Citation Data File: Lessons, Insights and Methodological Tools. NBER Working Paper 8498. National Bureau of Economic Research, Cambridge, MA. (Working-paper series carries no journal volume/issue; the paper has subsequent journal-handbook adaptations but the WP id is the canonical citation.)
PatentsView / U.S. Patent and Trademark Office. Patent bibliographic data via the USPTO Open Data Portal (https://api.uspto.gov; bulk tables at https://data.uspto.gov) — disambiguated patent, inventor, and assignee tables. (The standalone PatentSearch API at search.patentsview.org was paused on March 20, 2026 when PatentsView migrated to ODP; see data card uspto-patentsview.)

How this paper was built — annotated margin commentary

This section is not part of the paper. It is a build log for the next student, written to the camp’s reflective standard, explaining the choices behind the paper above so the template can be reused rather than copied.

Why this design, and where it sits in the course. The whole paper is the Week-4 event study (Chapter 4) pointed at a new kind of event — a patent grant instead of an earnings announcement — to build the Week-1 thing the field had despaired of: a measure of an unobservable (Chapter 6.1’s “measuring an idea”). I chose the KPSS construction because it is the clearest case in modern finance of converting a construct (the economic value of an invention) into a measure (a three-day price reaction), and because the Week-6 Reader’s Guide had already dissected its strengths and vulnerabilities, giving me a ready-made threats table to populate (Table 1). The paper structure follows Week-8 Chapter 8.3 exactly: title and abstract, an introduction that runs hook → one-sentence contribution → “I find” → roadmap, a positioning section, data, design with the identifying assumption stated in one sentence, results led by a stand-alone table, robustness, and a conclusion that ends on honest limits.

Why synthetic data, and why label it five times. Real CRSP returns are licensed and must stay on GMU infrastructure (CONVENTIONS §5), so a laptop-reproducible exemplar has to run on a synthetic stand-in. But the deeper reason is pedagogical: simulating the data lets me plant a known true patent value and then check whether the event study recovers it — something no empirical paper can do, because the truth is unobserved. That is the single most valuable thing this exercise teaches, so the synthetic nature is flagged in a banner at the top, in the abstract, in every table note, and twice in the conclusion. Over-labeling is deliberate: the failure mode I most wanted to prevent is a future reader lifting “t = 3.6” as if it described real firms. The numbers are all traceable to nb6.1 (60 firms, 633 patents, 217 firm-years, recovery 0.565/0.382/0.722, validation \hat\beta=0.0054/t=3.61, degradation 0.382→0.253 and t 3.61→2.45 across 0–75% contamination, top 10% holding 59% of value); I did not invent a single figure.

How the tables follow Appendix D. Table 2 is the headline, and it is built to stand alone (D.1 §6, D.2 §5): a reader covering the prose can still read off the dependent variable, the sample and its size, the estimator, the standard-error flavor, and what the t-column means, all from the note. I report the key coefficient (log innovation value) and one interpretable control regime, not a kitchen sink (D.2 §1). Crucially, I included the true-value benchmark row so the reader sees the ceiling and reads the estimated coefficient as attenuated by measurement error rather than as a clean effect — that contrast is the honest way to report a constructed measure. Standard errors are classical here and I say why in the note (firms are independent by construction in the synthetic panel), while flagging that real data would cluster by firm (CONVENTIONS §3); pretending to cluster on independent simulated data would be a false signal of rigor. I also kept significant digits matched and honest (D.1 §4): three decimals on the correlations, four on the small profitability coefficient because its standard error is uncertain in the fourth place, and no spurious precision dressed up to look authoritative.

Why the recovery correlation is deliberately imperfect. A student’s first instinct is to read 0.565 as a disappointing number and try to “fix” it. That instinct is exactly backwards, and the paper is structured to correct it. The median patent’s true value is a window blip well under 1%, while a firm’s three-day idiosyncratic return noise runs several percent — so a single window is, honestly, a noisy read, and a recovery correlation near 0.4–0.57 is the realistic signal-to-noise ratio, not a bug to engineer away. The pedagogical payoff is the jump to 0.722 once patents are aggregated to the firm-year: that is the whole reason KPSS-style measures are trusted at the firm-year level and distrusted as a single patent’s headline dollar figure, and showing both numbers side by side makes the lesson land harder than any caveat sentence could. I resisted the temptation to tune the noise floor down to produce a flattering correlation; the imperfection is the content.

Calibrating the verbs (D.5). The hardest discipline was language. The recovery and prediction results are predictive, not causal, so the verbs are “recovers,” “predicts,” “forecasts,” “is associated with” — never “causes” or “leads to.” The contribution sentence in §1 says what is now shown (that the method recovers a known truth and breaks under a known confound), not merely what I did, and its verb is capped by the weakest assumption. I ran the verb-audit D.5 prescribes — reading only the verbs attached to the headline result across all sections — to catch drift between the abstract and the conclusion, the place calibration usually fails. The conclusion’s honest-limits paragraph is the residual-concern column of Table 1 turned into prose: expectations-not-realized-value, efficient-markets circularity, selection, and the predictive-not-causal caveat each get one plain sentence.

What I would do with real data, and what I deliberately left out. The real pull is documented but not run: PatentsView (now served from the USPTO Open Data Portal, api.uspto.gov, post the March 20, 2026 cutover) for patents and grant dates, the NBER-lineage crosswalk and public KPSS file for the firm link and published values, CRSP for daily returns on GMU infrastructure with a pinned snapshot. On real data the first thing I would add is the placebo window as a live test rather than a by-construction pass (§6.2), and the most important extension is the systematic-confound case — contamination correlated with grant dates — which does not wash out under aggregation and is the genuine danger the random-contamination sweep cannot show. I left those as stated extensions rather than faking them, because the line between a result I produced and one I only described is exactly the line this paper is trying to teach a student to respect. The Hall–Jaffe–Trajtenberg working paper is cited at its canonical NBER WP id (no journal volume/issue exists for a WP); where any other citation could not be verified it is tagged [CHECK] per CONVENTIONS §6 — a visible flag of honesty beats a confident invention every time.

# Capstone 3 — Innovation from USPTO PatentsView > **ASSIP framing — read first.** *The ASSIP 2026 terminal deliverable is a research poster > (~48"×36") + a JSSR abstract (200–300 words) + an 8-minute poster talk + a replication packet > — not a full paper.* This exemplar is the **full research narrative the poster summarizes**: a > complete empirical-finance paper written one level above the ASSIP graded artifact, so that you > can see what the full argument looks like before you compress it onto six poster panels and a > 10-sentence abstract. Use it as a structural source for your ASSIP poster (each section maps to > a panel) and, if you choose to write the optional companion full paper, as a length-and-shape > template for that paper. > > **Illustrative results — read this first.** Every number in this paper comes from **synthetic patent and stock-return data constructed for instruction** in the Week-6 companion notebook `nb6.1` (`notebooks/week-06/nb6.1-patent-value-panel.ipynb`). The data are simulated from a seeded random generator with a *known* "true" patent value baked in, precisely so the construction can be checked against a ground truth no real dataset gives you. **These are not empirical findings.** No claim here describes the actual U.S. patent system, any real firm, or any real return. The paper is a gold-standard *template*: it shows what a finished innovation-measurement capstone looks like when it is built to the camp's standards (CONVENTIONS §4, Appendix D), with the real-data pull documented but not executed. Treat the magnitudes as worked-example outputs, not evidence. --- **Reading a Patent's Worth Off the Stock Price: A KPSS-Style Innovation Measure and a Grant-Date Event Study** *Leah R., ASSIP 2026 (Aspiring Scientists Summer Internship Program)* --- ## Abstract How much is a single invention worth? Patent counts treat a trivial design tweak and the transistor as one unit each; forward citations are slower and arrive a decade too late to value a recent patent. Following Kogan, Papanikolaou, Seru & Stoffman (2017), I build a market-based, dollar-denominated innovation measure: the abnormal stock return in a three-day window around a patent's grant date, scaled by market capitalization, read as the market's real-time estimate of the patent's private value. Working in a seeded **synthetic** panel of 60 firms and 633 patents over 2018–2021 — where the true patent value is known by construction — I reconstruct the measure with a Week-4 market-model event study and then run two tests. First, a *recovery* check: the reconstructed value correlates with the planted true value at 0.57 in levels and 0.38 in logs at the patent level, rising to 0.72 once aggregated to the firm-year, because averaging cancels the idiosyncratic window noise. Second, a *predictive* check: regressing next-year firm profitability on this year's log innovation value yields a positive, significant coefficient ($\hat\beta = 0.0054$, $t = 3.6$), the same validation pattern KPSS report on real data. A contamination experiment — injecting simulated earnings surprises into a rising share of grant windows — degrades both the recovery correlation (0.38 to 0.25) and the predictive $t$-statistic (3.6 to 2.5), making visible the confounding-news threat that the short window and firm-year aggregation are designed to blunt. The exercise demonstrates how to turn an unobservable (the economic value of an idea) into an observable, and exactly where the construction can break. --- ## 1. Introduction The single most cited patent of the 2010s and a patent that protects a slightly rounder bottle cap both count as "one patent." That is the problem that has dogged the economics of innovation since it began: the obvious way to measure how much a firm invents — count its patents — weights a breakthrough and a triviality equally, and the standard fix — count the forward citations a patent later attracts — is both slow, since citations accrue over ten or more years, and truncated, since a patent granted last year looks uncited only because the future has not happened yet. Neither proxy is denominated in the unit anyone making a capital-allocation decision actually cares about, which is dollars. An R&D manager, an investor, or a policymaker wants to know what an invention is *worth*, not how many of them there are. This paper builds and stress-tests a dollar-denominated innovation measure, in the spirit of Kogan, Papanikolaou, Seru & Stoffman (2017, hereafter KPSS), on a synthetic dataset where I know the right answer. **The contribution is pedagogical and methodological: I demonstrate, against a known ground truth, that a patent's private economic value can be read off the stock-market reaction to its grant, that the resulting firm-year measure predicts future profitability, and that confounding news in the grant window degrades the measure in a direction and magnitude I can quantify.** Because the data are simulated with a planted true value, I can do what no empirical paper can — compare the constructed measure to the truth it is trying to capture — which is the whole point of building it in a sandbox before trusting it in the wild. I find three things. First, the event-study reconstruction *recovers* the planted patent value: the correlation between estimated and true value is 0.57 in levels and 0.38 in logs across patents, and climbs to 0.72 when patents are aggregated to a firm-year panel, because firm-year sums average out the day-to-day idiosyncratic noise that contaminates any single three-day window. Second, the measure *validates* the way KPSS's does: regressing next-year firm profitability on this year's log innovation value gives a positive and statistically significant coefficient ($t = 3.6$), tracking the same predictive relationship the published measure exhibits on real firms. Third, the measure *breaks* in a legible way: when I contaminate a growing fraction of grant windows with simulated earnings-announcement returns, both the recovery correlation and the predictive $t$-statistic fall monotonically, exposing the confounding-window threat that is the central vulnerability of the design. The paper proceeds as follows. Section 2 positions the measure against the patent-counting and citation literature and states what KPSS added. Section 3 describes the synthetic data and the real-data pull it stands in for. Section 4 lays out the event-study design, states the identifying assumption in one sentence, and tabulates the threats to it. Section 5 reports the headline recovery and validation results. Section 6 runs the robustness checks — window length, a placebo window, and the contamination sweep. Section 7 concludes with the limits I cannot design away. ## 2. Positioning The literature on measuring innovation is a history of progressively better proxies, each fixing a defect in the last. The first generation counts **patents**: available far back in time, but treating every patent as equal, so a firm with a hundred trivial patents outscores the firm holding the one patent that defined a decade. The second generation, the workhorse of the field, counts **forward citations** — how many later patents cite this one — which genuinely tracks importance (the lineage runs through the NBER Patent Data Project of Hall, Jaffe & Trajtenberg). But citations are *slow*, accruing over a decade so a recent patent cannot be valued in real time, and *truncated*, since recent cohorts mechanically show fewer citations simply because less future has elapsed — making cross-cohort comparison the classic apples-to-oranges error. KPSS's move, which this paper imitates, is to let a third party that has already done the valuation work — the stock market — supply the number. When the U.S. Patent and Trademark Office grants a patent to a publicly traded firm, the grant is *news*: it resolves uncertainty about whether and on what claims the patent would issue. If the patent is valuable, the firm's stock should jump when the grant becomes public; if it is worthless, the stock should not move. KPSS therefore read the abnormal return in a short window around the grant date as the market's dollar-denominated estimate of the patent's private value, and released the resulting patent-level estimates publicly, which is why the measure has become standard infrastructure across finance, economics, and management. This is the same event-study logic Week 4 applied to corporate events; here the "event" is a patent grant, and the abnormal return *is* the valuation. My contribution relative to KPSS is not a new measure but a *transparent reconstruction*: by building it on data with a known answer, I show exactly how much signal the method recovers and how much it loses — which the real measure can never reveal, because the truth is unobserved. ## 3. Data The analysis runs on a **seeded synthetic panel** generated in `nb6.1` (random generator `np.random.default_rng(42)`), designed to mirror the schema of the real pull it stands in for. I simulate 60 firms over four calendar years, 2018–2021, at a daily frequency of roughly 252 trading days per year. Each firm has a market capitalization drawn on a log scale (median near \$2,000m, heavily right-skewed), a market beta near one, and an idiosyncratic daily volatility between 1.0% and 2.0%. Daily returns follow a market model, $R_{it} = a_i + b_i R_{mt} + \varepsilon_{it}$, the same normal-return model the event study will invert. Onto this return panel I graft **633 patents**, roughly two to three per firm-year, each carrying a grant date and a *true* dollar value drawn from a right-skewed log-normal distribution whose scale rises with a latent firm-quality term. The resulting value distribution is wildly skewed in exactly the way real patent values are: the top 10% of patents hold **59%** of the total planted value — the blockbuster phenomenon that is the entire motivation for not counting patents equally. Each patent's true value is then injected into the data as a small abnormal return spread over its three-day grant window, scaled so that a fixed dollar value is a large percentage move for a small firm and a tiny one for a giant (a \$50m patent moves a \$200m firm by 25% but a \$20,000m firm by 0.25%). This is the mechanism the event study must run in reverse. The synthetic panel deliberately reproduces the **selection** structure of the real data. Only patents whose firm has enough pre-grant return history to estimate a market model receive a value; the rest drop out, exactly as private firms, foreign assignees, and individual inventors drop out of the real measure for lack of a CRSP listing. After this screen, the patent-level estimates aggregate to a **217-row firm-year innovation-value panel**, the central data object. In a real build, the data would come from three stitched-together sources, documented in the data cards and in a non-executed (`RUN_REAL=False`) cell of `nb6.1`. Patents, grant dates, and citations would come from **USPTO PatentsView** via the **USPTO Open Data Portal** (`api.uspto.gov`, `X-API-KEY` header; PatentsView migrated to ODP on March 20, 2026), or the bulk PatentsView tables on ODP for panel-scale work. The patent-to-firm link and the published per-patent values would come from the **NBER-lineage crosswalk and the public KPSS patent-value file** (bulk downloads, no key). Daily firm returns would come from **CRSP**, which is licensed and must stay read-only on GMU infrastructure with a pinned snapshot date (CONVENTIONS §5). The join chain is PatentsView patent → crosswalk → `permno`/`gvkey` → CRSP/Compustat → firm outcomes. I name this pipeline so the template is reproducible, but I do not execute it, and no real number appears anywhere in this paper. ## 4. Empirical design ### 4.1 Constructing the measure The measure is built one patent at a time with a market-model event study, the Week-4 machine pointed at a patent grant. For patent $j$ owned by firm $i$ and granted on trading day $t_0$: 1. **Estimate the normal-return model.** On a pre-event estimation window of 120 trading days, ending 10 trading days before the grant (the gap prevents the event from leaking into the parameters), fit $R_{it} = \hat a_i + \hat b_i R_{mt}$ by least squares. 2. **Compute the abnormal return.** On each day $t$ of the three-day grant window (days $t_0$ through $t_0+2$), the abnormal return is $AR_{it} = R_{it} - (\hat a_i + \hat b_i R_{mt})$, the part of the return the market model did not predict. 3. **Cumulate and convert to dollars.** Sum the window abnormal returns into a cumulative abnormal return, $\mathrm{CAR}_{ij} = \sum_{t=t_0}^{t_0+2} AR_{it}$, and multiply by the firm's market capitalization to turn a percentage into a dollar value: $\widehat{V}_{ij} = \mathrm{CAR}_{ij}\times \text{MktCap}_i$. 4. **Aggregate.** Sum the (positive) per-patent dollar values within each firm and year to form the firm-year innovation-value panel. The one-line specification, in the camp's discipline (CONVENTIONS §4): **outcome** (constructed) = dollar value per patent; **key input** = idiosyncratic stock return in the three-day grant window; **adjustments** = market-model-filtered, scaled by market cap; **sample** = patents granted to firms with usable pre-grant return history; **the identifying assumption, in one sentence** = *the abnormal stock return in the three-day grant window equals the market's unbiased estimate of the patent's economic value, because markets impound the patent's value at grant and no other value-relevant news contaminates the window.* For the **predictive** validation, the specification is a firm-year regression: **outcome** = next-year firm profitability; **key regressor** = log of this year's innovation value; **controls** = log market capitalization; **fixed effects** = year; **standard errors** = classical (the synthetic panel is independent across firms by construction, so there is no within-firm clustering to correct; on real data this regression would cluster by firm, per CONVENTIONS §3); **sample** = the 217-row firm-year panel. This relationship is **predictive, not causal** — I do not claim innovation *causes* profits, only that the measure forecasts them the way a valid value measure should. ### 4.2 Threats to identification The identifying assumption has three soft joints, each inherited directly from KPSS's "what's vulnerable." Table 1 names each threat, its direction, and the design response. **Table 1 — Threats to the grant-window value measure, and responses** | Threat | Why it biases the measure | Direction | Design response | |---|---|---|---| | **Confounding window news** (earnings, deals, analyst revisions landing in the 3-day window) | The CAR mixes the patent's value with unrelated news, so $\widehat{V}$ no longer isolates the patent | Either sign per patent; attenuates the measure on average if idiosyncratic | Short (3-day) window limits exposure; firm-year aggregation averages out *idiosyncratic* confounds; quantified in §6.3 | | **Selection to granted patents of public firms** | Only granted patents of firms with return history get a value; rejected/abandoned patents and private, foreign, university assignees are invisible | Bounds the population, not the within-sample estimate | Disclosed, not corrected; restricts external validity (stated in §7) | | **Expectations, not realized value** | The window return is the market's *forecast* at grant, not what the patent ultimately earned; bias if the market is systematically optimistic about flashy patents | Inflates value of patents the market over-rates | None available within the design; the measure is "market's expected value at grant," and prose says so | | **Efficient-markets assumption baked in** | The measure assumes prices impound value promptly; if markets are slow the window misses leaked value | Attenuates | Short window plus sensitivity to window length (§6.1); cannot test efficiency with a measure that assumes it | The first threat is the one I can manipulate in a synthetic world, so it is where §6.3 aims its experiment. The remaining three I disclose and bound rather than solve, which is the honest posture for a measurement design. ## 5. Results The headline result is the validation table: does the measure, reconstructed from three-day stock blips, both recover the planted truth and predict future outcomes? Table 2 reports both checks, with the true-value regression as a benchmark ceiling. **Table 2 — Recovery of the planted value and prediction of next-year profitability** | | Recovery: corr(true, estimated) | | Prediction: next-year profitability on log innovation value | | | |---|---:|---:|---:|---:|---:| | | Level | Log | $\hat\beta$ | SE | $t$ | | **Estimated measure** (event-study CAR × mktcap) | 0.565 | 0.382 | 0.0054 | 0.0015 | 3.61 | | Firm-year aggregate (estimated) | 0.722 | — | | | | | *Benchmark: true planted value* | *1.000* | *1.000* | *0.0152* | *0.0017* | *9.20* | *Notes.* Synthetic data from `nb6.1`; **illustrative, not empirical**. 633 patents and a 217-row firm-year panel, 60 firms × 2018–2021. "Recovery" columns report the Pearson correlation between the event-study-estimated patent value and the known planted true value, in levels and in logs (logs computed only where the estimated value is positive, since a CAR can come out negative). The firm-year row aggregates patent values within firm-year before correlating. The prediction columns report a firm-year OLS of next-year profitability on log innovation value, controlling for log market cap and year fixed effects; the coefficient on log innovation value is shown. Standard errors are classical (firms are independent by construction in the synthetic panel; a real-data version would cluster by firm). The benchmark row substitutes the *true* planted value for the estimated one, giving the ceiling the estimate would hit with zero measurement error. $^{\*\*\*}\,p<0.01$. Coefficient $0.0054$ has $t=3.61$, $p=0.0004$. Three readings. First, **the event study recovers the truth, imperfectly but unmistakably.** The patent-level correlation of 0.565 in levels (0.382 in logs) is far from perfect, and it *should* be: the median patent's true value is a window blip of well under 1%, while a firm's three-day idiosyncratic noise runs several percent, so any single window is a noisy read. This is not a flaw to apologize for; it is the realistic signal-to-noise ratio that *motivates* the next number. Second, **aggregation is the cure for window noise.** Summing patents to the firm-year lifts the recovery correlation to 0.722, because the idiosyncratic noise in individual windows is mean-zero and cancels when many patents are pooled, while the signal accumulates. This is the mechanical reason KPSS-style measures are most trustworthy at the firm-year level — the level at which the economics is actually done — and least trustworthy as a single patent's headline dollar figure. Third, **the measure predicts the future the way a valid value measure should.** The estimated log innovation value carries a positive, significant coefficient on next-year profitability ($\hat\beta = 0.0054$, $t = 3.61$). The benchmark row shows the ceiling: with the *true* value substituted in, the coefficient is larger ($0.0152$) and the $t$-statistic far higher ($9.20$). The gap between the two rows is precisely the cost of measurement error — the estimated measure attenuates the relationship toward zero, exactly as classical measurement error in a regressor predicts, but does not destroy it. A three-day stock reaction, filtered and scaled, forecasts a real outcome a year out. ## 6. Robustness A measurement design earns trust by surviving perturbations of the choices that built it. Three checks: the window length, a placebo window, and the confounding-news experiment that is the point of working in a sandbox. ### 6.1 Window length The three-day window is a choice, and the right way to disclose a choice is to vary it. Shortening to a one-day window $(0,0)$ throws away abnormal return that lands on days $t_0+1$ and $t_0+2$, attenuating the measure; lengthening to a five-day window $(0,4)$ admits more unrelated news, raising the variance of the estimate and inviting the confounding-window threat of §6.3. The three-day window is the standard compromise — long enough to capture a reaction that does not all land on the grant date, short enough to keep the confound window narrow. The qualitative lesson, which `nb6.1`'s *Your Turn* extension makes a student reproduce, is that the recovery correlation is *not invariant* to the window; reporting the measure without disclosing the window would hide a real degree of freedom. ### 6.2 Placebo window The cleanest falsification is a placebo: run the identical event study on a window where *no* patent value was injected — a stretch of trading days unconnected to any grant. By construction the synthetic data plant value only in true grant windows, so a placebo window should return a CAR that is mean-zero noise and an estimated value indistinguishable from zero, with no correlation to any patent's planted value. A placebo that instead produced apparent "value" would signal that the event-study machine was manufacturing signal out of ordinary volatility — the single most important thing to rule out before trusting the real-window estimates. The design passes this check by construction here; on real data it is the test that separates a genuine grant-window reaction from a mechanical artifact, and it is the one I would run first. ### 6.3 Contaminating the grant windows The headline robustness exercise attacks the design at its weakest joint (Table 1, row 1): confounding news in the window. I inject a simulated one-day earnings surprise — a normal shock with a 6% standard deviation, large relative to the patent's own window signal — into a rising fraction of grant windows, rebuild the measure on the contaminated data, and watch both the recovery correlation and the predictive $t$-statistic degrade. Results are in Table 3, averaged over five contamination seeds for smoothness. **Table 3 — Confounding window news degrades both recovery and prediction** | Share of windows contaminated | 0% | 15% | 30% | 50% | 75% | |---|---:|---:|---:|---:|---:| | Recovery correlation (log) | 0.382 | 0.353 | 0.330 | 0.291 | 0.253 | | Validation $t$ (next-year profit) | 3.61 | 3.40 | 3.13 | 2.81 | 2.45 | *Notes.* Synthetic data from `nb6.1`; **illustrative, not empirical**. A one-day earnings surprise (Gaussian, SD = 6%) is added to the first day of a random fraction of grant windows; the value measure is then rebuilt from the contaminated returns. "Recovery correlation (log)" is corr(true, estimated) at the patent level; "Validation $t$" is the $t$-statistic on log innovation value in the next-year-profitability regression of Table 2. Each column averages five contamination seeds. The 0% column reproduces the clean estimates of Table 2. The pattern is monotone and interpretable. As contamination rises from 0% to 75%, the recovery correlation falls from 0.382 to 0.253 and the predictive $t$-statistic falls from 3.61 to 2.45 — the measure is being eaten by noise the event study cannot distinguish from the patent's own signal. Two lessons follow. First, this is *why* the short window and firm-year aggregation are the defenses: a narrow window gives confounds less room to land, and aggregation averages out *idiosyncratic* confounds like a one-off earnings surprise. Second, and this is the subtler point KPSS's "what's vulnerable" insists on, aggregation only saves the measure when the confound is idiosyncratic. The dangerous case — flagged in `nb6.1`'s extension and not run here — is *systematic* contamination, where good-news days correlate with grant dates; that bias does not wash out under aggregation and would survive into the firm-year panel. ## 7. Conclusion This paper reconstructs a KPSS-style, dollar-denominated innovation measure on synthetic data with a known answer, and uses that answer to show three things the real measure can never directly show: the event study *recovers* the planted patent value (correlation 0.57 at the patent level, 0.72 once aggregated to the firm-year), the measure *predicts* next-year profitability ($t = 3.6$, against a true-value ceiling of $t = 9.2$), and confounding window news *degrades* both recovery and prediction in a monotone, quantifiable way. The exercise turns an unobservable — the economic value of an idea — into an observable, and it locates the exact seams where the construction can fail. The limits are real and I state them plainly. **The measure is the market's *expected* value at grant, not realized value;** a patent the market loved at grant and that later flopped still scores high, and any sentence dropping the words "expected" and "the market's" overclaims. **The measure assumes efficient markets,** so it cannot be used to *test* efficiency without circularity, and it is sensitive to the choice of normal-return model and window length (§6.1). **The sample is selected** to granted patents of firms with usable return history; the real-data analogue excludes private firms, foreign assignees, universities, and de-listed firms, so the firm-year panel describes the innovation of large, listed, surviving firms and nothing I show here generalizes to the whole economy. And the **predictive** validation is an association, not a causal effect — innovating firms grow and the measure forecasts it, but I have no random assignment of patents, no instrument, no discontinuity, and reverse causality (growing firms patent more) and common drivers (a demand shock lifting both) remain live alternatives. Above all, **the data are synthetic.** The numbers demonstrate that the *method* recovers a known truth and breaks under a known confound; they say nothing about real patents, real firms, or real returns. That separation — between validating a method and discovering a fact — is the discipline the whole exercise is meant to teach. --- ## References - Kogan, L., Papanikolaou, D., Seru, A., & Stoffman, N. (2017). Technological Innovation, Resource Allocation, and Growth. *Quarterly Journal of Economics*, 132(2), 665–712. - Hall, B. H., Jaffe, A. B., & Trajtenberg, M. (2001). The NBER Patent Citation Data File: Lessons, Insights and Methodological Tools. *NBER Working Paper 8498*. National Bureau of Economic Research, Cambridge, MA. (Working-paper series carries no journal volume/issue; the paper has subsequent journal-handbook adaptations but the WP id is the canonical citation.) - PatentsView / U.S. Patent and Trademark Office. Patent bibliographic data via the **USPTO Open Data Portal** (`https://api.uspto.gov`; bulk tables at `https://data.uspto.gov`) — disambiguated patent, inventor, and assignee tables. (The standalone PatentSearch API at `search.patentsview.org` was paused on March 20, 2026 when PatentsView migrated to ODP; see data card `uspto-patentsview`.) --- --- ## How this paper was built — annotated margin commentary *This section is not part of the paper. It is a build log for the next student, written to the camp's reflective standard, explaining the choices behind the paper above so the template can be reused rather than copied.* **Why this design, and where it sits in the course.** The whole paper is the Week-4 event study (Chapter 4) pointed at a new kind of event — a patent grant instead of an earnings announcement — to build the Week-1 thing the field had despaired of: a *measure* of an unobservable (Chapter 6.1's "measuring an idea"). I chose the KPSS construction because it is the clearest case in modern finance of converting a construct (the economic value of an invention) into a measure (a three-day price reaction), and because the Week-6 Reader's Guide had already dissected its strengths and vulnerabilities, giving me a ready-made threats table to populate (Table 1). The paper structure follows Week-8 Chapter 8.3 exactly: title and abstract, an introduction that runs hook → one-sentence contribution → "I find" → roadmap, a positioning section, data, design with the identifying assumption stated in one sentence, results led by a stand-alone table, robustness, and a conclusion that ends on honest limits. **Why synthetic data, and why label it five times.** Real CRSP returns are licensed and must stay on GMU infrastructure (CONVENTIONS §5), so a laptop-reproducible exemplar has to run on a synthetic stand-in. But the deeper reason is pedagogical: simulating the data lets me *plant a known true patent value* and then check whether the event study recovers it — something no empirical paper can do, because the truth is unobserved. That is the single most valuable thing this exercise teaches, so the synthetic nature is flagged in a banner at the top, in the abstract, in every table note, and twice in the conclusion. Over-labeling is deliberate: the failure mode I most wanted to prevent is a future reader lifting "$t = 3.6$" as if it described real firms. The numbers are all traceable to `nb6.1` (60 firms, 633 patents, 217 firm-years, recovery 0.565/0.382/0.722, validation $\hat\beta=0.0054$/$t=3.61$, degradation 0.382→0.253 and $t$ 3.61→2.45 across 0–75% contamination, top 10% holding 59% of value); I did not invent a single figure. **How the tables follow Appendix D.** Table 2 is the headline, and it is built to stand alone (D.1 §6, D.2 §5): a reader covering the prose can still read off the dependent variable, the sample and its size, the estimator, the standard-error flavor, and what the $t$-column means, all from the note. I report the *key coefficient* (log innovation value) and one interpretable control regime, not a kitchen sink (D.2 §1). Crucially, I included the **true-value benchmark row** so the reader sees the ceiling and reads the estimated coefficient as *attenuated by measurement error* rather than as a clean effect — that contrast is the honest way to report a constructed measure. Standard errors are classical here and I say *why* in the note (firms are independent by construction in the synthetic panel), while flagging that real data would cluster by firm (CONVENTIONS §3); pretending to cluster on independent simulated data would be a false signal of rigor. I also kept significant digits matched and honest (D.1 §4): three decimals on the correlations, four on the small profitability coefficient because its standard error is uncertain in the fourth place, and no spurious precision dressed up to look authoritative. **Why the recovery correlation is deliberately imperfect.** A student's first instinct is to read 0.565 as a disappointing number and try to "fix" it. That instinct is exactly backwards, and the paper is structured to correct it. The median patent's true value is a window blip well under 1%, while a firm's three-day idiosyncratic return noise runs several percent — so a single window is, honestly, a noisy read, and a recovery correlation near 0.4–0.57 is the *realistic* signal-to-noise ratio, not a bug to engineer away. The pedagogical payoff is the jump to 0.722 once patents are aggregated to the firm-year: that is the whole reason KPSS-style measures are trusted at the firm-year level and distrusted as a single patent's headline dollar figure, and showing both numbers side by side makes the lesson land harder than any caveat sentence could. I resisted the temptation to tune the noise floor down to produce a flattering correlation; the imperfection *is* the content. **Calibrating the verbs (D.5).** The hardest discipline was language. The recovery and prediction results are *predictive*, not causal, so the verbs are "recovers," "predicts," "forecasts," "is associated with" — never "causes" or "leads to." The contribution sentence in §1 says what is now *shown* (that the method recovers a known truth and breaks under a known confound), not merely what I *did*, and its verb is capped by the weakest assumption. I ran the verb-audit D.5 prescribes — reading only the verbs attached to the headline result across all sections — to catch drift between the abstract and the conclusion, the place calibration usually fails. The conclusion's honest-limits paragraph is the residual-concern column of Table 1 turned into prose: expectations-not-realized-value, efficient-markets circularity, selection, and the predictive-not-causal caveat each get one plain sentence. **What I would do with real data, and what I deliberately left out.** The real pull is documented but not run: PatentsView (now served from the USPTO Open Data Portal, `api.uspto.gov`, post the March 20, 2026 cutover) for patents and grant dates, the NBER-lineage crosswalk and public KPSS file for the firm link and published values, CRSP for daily returns on GMU infrastructure with a pinned snapshot. On real data the first thing I would add is the **placebo window** as a live test rather than a by-construction pass (§6.2), and the most important extension is the **systematic-confound** case — contamination correlated with grant dates — which does *not* wash out under aggregation and is the genuine danger the random-contamination sweep cannot show. I left those as stated extensions rather than faking them, because the line between a result I produced and one I only described is exactly the line this paper is trying to teach a student to respect. The Hall–Jaffe–Trajtenberg working paper is cited at its canonical NBER WP id (no journal volume/issue exists for a WP); where any other citation could not be verified it is tagged `[CHECK]` per CONVENTIONS §6 — a visible flag of honesty beats a confident invention every time.