Appendix — Solutions — PS 6.1 (Building the KPSS Patent-Value Measure: Construction, Validation, and Critique)

Problem set: book/weeks/week-06/ps6.1.md (PS 6.1, Week 6). Chapter: Ch 6.1 — Reader’s Guide: Technological Innovation, Resource Allocation, and Growth (Kogan, Papanikolaou, Seru & Stoffman 2017), read through the seven-part Reader’s-Guide anatomy.

These are full worked solutions, not just answers. Notation follows CONVENTIONS.md and locks to Ch 6.1 and the prior weeks it builds on: the market model R_{it}=a_i+b_iR_{mt}+\varepsilon_{it} fit on a pre-event estimation window; the day-by-day abnormal return AR_{ft_k}=R_{ft_k}-(\hat a_f+\hat b_f R_{mt_k}); the cumulative abnormal return \text{CAR}=\sum_k AR_{ft_k} over the short grant window; the dollar conversion by market capitalization; the §2 three-refinement construction (filter the market → scale by market cap → inflate by 1/\pi for grant-anticipation) giving the per-patent value \xi_j; the Week-1 construct-vs-measure gap, convergent/predictive validity, selection, and external validity; and the variance-of-an-average shrinkage \sigma\mapsto\sigma/\sqrt n. Every numerical result here was confirmed in Python (NumPy abnormal-return and CAR arithmetic, the three-window stress, and the 1/\sqrt n table); verifying notes appear where useful. All dollar figures and table values are illustrative and constructed, never quoted as KPSS’s own; the paper is cited by name as Kogan, Papanikolaou, Seru & Stoffman (2017) (KPSS).

Problem 1 — Why a dollar measure? Counts, citations, and the construct–measure gap (14 points)

(a) (4 pts) The construct Leah cares about is the private economic value of a single invention — how many dollars the patent is worth to the firm that owns it. The two standard proxies fail in distinct ways:

Patent counts weight every patent equally — one trivial design tweak counts the same as the patent that defined a decade — so a firm with a hundred throwaway patents looks more innovative than a firm with the one transformative one. Counts measure quantity, not value.

Forward citations are a real improvement (they track importance), but they carry two defects §1 insists you recite. First, timing/speed: citations accumulate over ten or more years, so you cannot value a recent patent in real time. Second, truncation: recent patents look artificially uncited simply because the future that would cite them has not happened yet. And neither proxy is denominated in dollars — the unit anyone actually cares about.

(b) (4 pts) The clever move (§5) is to let the stock market value the patent for you. (i) A patent grant is news because the USPTO’s decision and exact claims were uncertain beforehand; the grant resolves some of that uncertainty, and the market updates the firm’s price the moment the grant becomes public. (ii) A large positive reaction says the market judged the patent valuable; a near-zero reaction says it judged it worthless — the size of the jump is the market’s dollar estimate. (iii) This outsources the valuation to the thousands of investors who collectively set the price and have already done the hard work, so KPSS need not build a structural model of demand, competition, and discount rates (assumptions nobody can defend); they just read the price reaction with the cleanest tool available, the abnormal return in a tight window.

(c) (3 pts) The units advantage: citations are not denominated in money — they count attention, not dollars — whereas the market reaction is a dollar figure directly. The speed/timing advantage: a stock reaction happens in three days and is available the instant the patent issues, while citations take a decade to accrue (and are truncated for recent patents). The speed is what makes a long firm-year panel possible: because every patent — including last year’s — gets a value immediately, KPSS can build a dense year-by-year panel, and a panel is exactly what you need to study growth and reallocation, which play out over time. Citations could never deliver a real-time panel.

(d) (3 pts) In Week-1 language, the construct is the patent’s true economic value (unobservable), and the measure is the grant-window abnormal stock return, scaled to dollars (observable). The honest one-sentence description: the KPSS number is the market’s expected private value of the patent at grant. Dropping any of the three words overclaims: drop “market’s” and you pretend it is the true value rather than the market’s opinion of it; drop “expected” and you pretend it is realized value rather than a forecast; drop “at grant” and you pretend it incorporates everything learned later. It is not the value the patent ultimately realized — a patent the market loved at grant that later flopped still carries a high KPSS value, which is precisely the construct-vs-measure gap.

Problem 2 — Rebuild one patent’s value from a grant-window event study (24 points)

Given: \hat a_f=0.0000, \hat b_f=1.10, \text{MktCap}_f=\$8{,}000{,}000{,}000. Predicted normal return each day is \hat R_{ft}=\hat a_f+\hat b_f R_{mt}=1.10\,R_{mt}.

(a) (6 pts) Abnormal return AR_{ft_k}=R_{ft_k}-(\hat a_f+\hat b_f R_{mt_k}), day by day:

k	R_{ft}	R_{mt}	\hat b_f R_{mt}	AR_{ft_k}=R_{ft}-\hat b_f R_{mt}
-1	0.0120	0.0040	1.10(0.0040)=0.0044	0.0120-0.0044=0.0076 (0.76\%)
0	0.0180	0.0050	1.10(0.0050)=0.0055	0.0180-0.0055=0.0125 (1.25\%)
+1	0.0070	0.0030	1.10(0.0030)=0.0033	0.0070-0.0033=0.0037 (0.37\%)

(With \hat a_f=0 the intercept drops out.) You subtract \hat b_f R_{mt} because of the §2 first refinement, “filter out the market”: a stock can move during the window for reasons that have nothing to do with the patent — the whole market rallied, or the firm’s sector moved — and that contribution lives in \hat b_f R_{mt}. If you skipped it, the raw window return would mix the patent’s value with whatever the broad market did those three days; the patent’s value lives in the residual (the abnormal return), not the raw return.

(b) (5 pts) Cumulative abnormal return over the three window days: \text{CAR}=\sum_{k=-1}^{+1}AR_{ft_k}=0.0076+0.0125+0.0037=0.0238=2.38\%. Gross dollar reaction: \text{CAR}\times\text{MktCap}_f=0.0238\times\$8{,}000{,}000{,}000=\$190{,}400{,}000. The market-cap multiplication is §2’s second refinement, “scale by market cap.” A raw percentage cannot be compared across firms of different sizes because the same 2.38% jump is a very different dollar amount for an $8B firm than for a $200M firm; multiplying by market cap converts the firm-specific percentage surprise into the dollars the market added to the firm’s value on news of the patent.

(c) (5 pts) The third refinement inflates for grant-anticipation. With \pi=0.5: \xi_j=\frac{\text{CAR}\times\text{MktCap}_f}{\pi}=\frac{\$190{,}400{,}000}{0.5}=\$380{,}800{,}000. The adjustment raises the estimate because a grant is only partly a surprise — markets anticipate some grants, so the observed window reaction captures only the unanticipated fraction of the patent’s value; dividing by \pi<1 grosses the reaction back up to the full expected value. Interpretation, the careful sentence now attached to a number: \xi_j\approx\$381 million is the market’s expected private value, at grant, of patent j to firm f — not what the patent ultimately earned.

(d) (5 pts) Window stress. Holding \hat a_f,\hat b_f fixed and converting gross (pre-\pi) reactions to dollars:

One-day (k=0 only): \text{CAR}_1=AR_{f,0}=1.25\%=0.0125, so 0.0125\times\$8\text{B}=\$100{,}000{,}000.
Three-day (k\in\{-1,0,+1\}): \text{CAR}_3=2.38\%, so \$190{,}400{,}000 (from part b).
Five-day (k\in\{-2,\dots,+2\}): add AR_{f,-2}=+0.005\%=0.00005 and AR_{f,+2}=-0.15\%=-0.0015 to the three-day CAR: \text{CAR}_5=0.0238+0.00005-0.0015=0.02235=2.235\%, so 0.02235\times\$8\text{B}=\$178{,}800{,}000.

Side by side: $100M (1-day) · $190.4M (3-day) · $178.8M (5-day) — the estimate nearly doubles between the one-day and three-day windows and then drifts back down when two more days are added. This makes §6’s vulnerability “the measure depends on modeling choices” concrete: the same patent’s “value” moves by tens of millions of dollars purely because of the window you pick. The short window is the principled default because the longer you make the window, the more time you give other value-relevant news (earnings, deals, analyst revisions) to land inside it and contaminate the abnormal return — the very confound of §6 / Problem 4(b). So no single \xi_j should be read as “the patent is worth exactly $X”; it is a noisy, model-dependent point estimate, which is exactly why the firm-year aggregation of Problem 5 is the unit the paper actually trusts.

(Verification: AR = \{0.76,1.25,0.37\}\%; CAR_3=2.38%, $190.4M; CAR_1=1.25%, $100.0M; CAR_5=2.235%, $178.8M; \xi_j=$380.8M at \pi=0.5. Confirmed in Python.)

Problem 3 — What would convince you the measure is valid? (18 points)

(a) (5 pts) Convergent validity. This is §4 Step 1’s first sanity test — “does the market-based value correlate with forward citations?” You wanted a positive, significant slope, and Leah’s +0.31 (t=5.8) delivers it. The logic of why it is reassuring but not sufficient: citations are the literature’s established proxy for patent importance, trustworthy where they apply, so if the dollar value rises with citations the new measure agrees with the old one where the old one is trustworthy — that is convergent validity, evidence the new number is tracking something real and not noise. But it is not sufficient because a measure that merely reproduced citations would add nothing — citations already exist, and they are slow and unit-less; the point of KPSS is to do more than citations (dollars, in real time), so convergent validity is the floor, not the ceiling. (Acceptable also: citations themselves are an imperfect proxy, so agreement with them cannot by itself certify agreement with the construct.)

(b) (5 pts) Predictive validity. §4 calls the profitability/growth test “the strongest” because it shows the measure — built from a three-day stock reaction — forecasts a real outcome years out, the owning firm’s future profits/output. In Week-1 terms: the citation test in (a) shows the measure agrees with another proxy (one imperfect ruler matching another), whereas the profitability test shows the measure agrees with the construct itself — economic value should, if real, be followed by higher realized profits, so a positive predictive slope is direct evidence that the number captures value, not just attention. Leah’s t\approx3.6 (predicting a planted next-year outcome in nb6.1) is exactly this stronger evidence: a forecast of reality, not a correlation with a sibling proxy.

(c) (4 pts) The discriminant check. The §4 test is “does it survive the obvious confound?” — showing the patent-grant signal is distinct from a generic announcement return rather than just reflecting other contemporaneous news. One concrete check on synthetic data: Leah compares the average grant-window abnormal return to the average abnormal return over matched non-grant windows for the same firms (placebo windows on ordinary days with no patent news) — if grant windows show a systematically larger reaction than placebo windows, the signal is attributable to the patent rather than to generic announcement-day churn. (Acceptable also: compare across patents that do vs do not later prove valuable, or regress out earnings-announcement returns and check the grant effect survives.)

(d) (4 pts) Falsifiability. §4’s rule for reading these tables is to read them for signs and significance, not for magnitudes you memorize — the pattern that matters is that the value measure is positively and significantly related to citations, profits, and growth. The measure would break (“the paper collapses here”) if that pattern failed: if the value–citation slope were flat or wrong-signed, if the value–profit forecast were insignificant, or if grant-window returns were indistinguishable from generic announcement returns — any of these would mean the constructed dollar value is not tracking patent value at all. This entire validation exercise lives in Reader’s-Guide section §4, “Table-by-table reading order” (specifically Step 1, the measure-validation evidence a referee reads first).

Problem 4 — Read like a referee: three named vulnerabilities (24 points)

(a) (8 pts) Expectations, not realized value (Week 1 — measurement validity).

The grant-window return is a forecast of the patent’s economic value, formed by the market in three days using only the information available then. It is not the value the patent ultimately realized. The gap is exactly the Week-1 construct-vs-measure gap: the construct is true realized economic value; the measure is a three-day price reaction, i.e., the market’s expectation of that value at grant.
Concrete example, Leah’s terms: a glamorous AI-chip patent grant that the market greets with a big positive abnormal return — earning a high \xi_j — but whose technology is leapfrogged within two years and never earns the firm a dollar. For “flashy” patents the market over-loves at grant, the measure’s error is upward: \xi_j is too high relative to realized value (and symmetrically, quiet patents the market underrates carry a downward-biased \xi_j).
The check (from the “Read like a referee” box): regress a later realized outcome — eventual profits attributable to the patent, eventual forward citations, or litigation/renewal outcomes — on the grant-window \xi_j, and ask whether the forecast is unbiased (slope \approx1 on average, errors centered at zero) or systematically too high for some kinds of patents (e.g., a negative residual concentrated among the flashiest, most-hyped technology classes). The deep reason the measure can never, by itself, separate “the market was right” from “the market was wrong but consistently wrong”: the measure is the market’s forecast, so any test using only the measure is circular — you need an external realized-outcome benchmark (the later data above) to grade the forecast, and even then a consistent bias would just shift the whole calibration without revealing itself within the grant-window number alone.

(b) (8 pts) Confounding news in the window (Week 4 — event-study identification).

The identifying assumption (§2, one sentence): the abnormal return in the grant window equals the market’s unbiased estimate of the patent’s value — as long as the patent grant is the only value-relevant news in that short window (and markets are efficient enough to impound it).
Two concrete other-news kinds: an earnings announcement (or a guidance update) landing in the window, and a merger/acquisition or major-deal announcement (also acceptable: an analyst rating revision, a lawsuit outcome, a regulatory ruling). Either corrupts \xi_j because the window’s abnormal return now mixes the patent’s value with the other news’s value — the price moved for two reasons and the construction credits all of it to the patent, biasing \xi_j up or down by the size of the contaminating surprise.
Idiosyncratic confounds are news items that hit individual patents in random, uncorrelated directions — some windows catch good news, some bad, with no relationship to the patent’s true value. Averaging over the many patents in a firm-year (Problem 5) makes these cancel, because their mean tends to zero. Systematic confounds are contamination that is correlated across patents or correlated with the true value itself — e.g., if firms strategically time good-news disclosures near grant dates, or if grant windows for genuinely valuable patents also tend to carry positive earnings surprises. These do not wash out under aggregation, because their mean does not go to zero — averaging many same-signed errors leaves the bias intact. The systematic case is the genuinely dangerous one, and it is exactly the nb6.1 “break it” extension where injecting contamination correlated with the true value survives aggregation and degrades the validation (whereas purely random contamination is averaged away even at high contamination shares).

(c) (8 pts) Selection and efficient markets (Week 1 — external validity; circularity).

The three stacked selection screens (§6): (1) only granted patents get a value — rejected/abandoned applications, and the decision to even apply, are invisible; (2) only patents of publicly traded firms with usable CRSP returns can be valued — private firms, pre-IPO startups, universities, and foreign assignees drop out; (3) the firm must survive and stay listed around the grant date. The population the firm-year panel actually represents: the innovation of large, public, surviving U.S. firms — not “all innovation in the economy.”
A conclusion not extendable to the whole economy: that a surge of innovation reallocates resources away from rivals across the economy (the creative-destruction story at the aggregate level). The panel only sees public, surviving firms, so it is silent on innovation by private firms, startups, and universities — exactly the actors who do much of the disruptive innovating — so an economy-wide reallocation claim requires an argument the panel’s selected sample cannot supply.
The circularity: the measure is built on the assumption that the market prices the patent correctly and promptly at grant — efficiency is the premise that turns the abnormal return into “the value.” So you cannot turn around and use the measure to test market efficiency, because you would be assuming the conclusion. “Efficient markets cuts both ways” for the measure’s accuracy: if markets are slow/noisy, value leaks in after the window and the measure is attenuated (too low); if markets over-react, the window over-counts and the measure is inflated (too high). Either way the measure inherits whatever pricing errors the market makes, and the construction has no internal way to detect them.

Problem 5 — Aggregation: why summing noise sharpens the signal (10 points)

(a) (5 pts) Write \hat\xi_j=\xi_j^{\text{true}}+e_j with e_j roughly independent across patents and \operatorname{Var}(e_j)=\sigma^2. For the n patents a firm is granted in a year, the average noise is \bar e=\frac1n\sum_{j=1}^n e_j. By the variance-of-an-average rule, with independent e_j, \operatorname{Var}(\bar e)=\frac{1}{n^2}\sum_{j=1}^n\operatorname{Var}(e_j)=\frac{n\sigma^2}{n^2}=\frac{\sigma^2}{n} \;\Longrightarrow\; \operatorname{sd}(\bar e)=\frac{\sigma}{\sqrt n}. So averaging n roughly-independent noisy values shrinks the noise standard deviation by the factor 1/\sqrt n:

Patents in firm-year, n	1	4	25	100
Noise-SD shrink factor 1/\sqrt{n}	1.00	0.50	0.20	0.10

(A firm with 100 granted patents averages its per-patent noise down to one-tenth of a single patent’s noise SD. Equivalently, the firm-year total has signal-to-noise growing like \sqrt n: the true total scales with n while its noise SD scales only with \sqrt n\,\sigma, so SNR \sim n/(\sqrt n\,\sigma)=\sqrt n/\sigma.)

(b) (3 pts) The shrinkage requires the noise e_j to be roughly independent (uncorrelated) across patents with mean zero — only then do the cross terms in the variance vanish and the errors cancel rather than reinforce. This is the same distinction as Problem 4(b): aggregation defeats idiosyncratic confounds (independent, mean-zero window noise — which is what makes \bar e\to0), and it leaves systematic confounds untouched (correlated or value-dependent contamination, whose mean does not go to zero, so \bar e stays biased). It is the same fork because whether the firm-year panel can be trusted comes down to exactly whether the per-patent errors are independent (washed out by summing) or systematic (carried straight through the sum).

(c) (2 pts) Leah’s firm-year correlation (\approx0.72) beats her per-patent correlation (\approx0.4) because the construction is unchanged but the unit changed: at the patent level each \hat\xi_j carries the full per-patent noise SD \sigma (window confounds, model error swamping a small true blip), whereas summing within firm-year averages that noise down by 1/\sqrt n, so the firm-year totals track the true firm-year value far more tightly. The design lesson for her capstone: when a per-unit measure is noisy but unbiased, aggregating to a coarser unit buys you a much cleaner signal (noise falls like 1/\sqrt n while the true signal accumulates), at the cost of resolution — the firm-year panel can no longer answer patent-level questions (which specific patent was the blockbuster), only firm-year ones.

Problem 6 — Design the replication pipeline for `nb6.1` (10 points)

(a) (4 pts) Inspect and aggregate (§7 Exercise 1). Ordered steps: 1. Load the patent-level value file (one row per granted patent: firm id, grant year, \xi_j), or the seeded synthetic stand-in with the same schema. 2. Plot the distribution of per-patent values first, before trusting them — expect it to be wildly right-skewed, a few blockbuster patents dwarfing thousands of small ones (in nb6.1’s synthetic data the top decile holds the large majority of total value). That skew is the whole motivation for not counting patents equally: if value were roughly uniform across patents, counts would suffice; because it is heavy-tailed, one transformative patent outweighs a hundred trivial ones, so you must weight by dollars. 3. Aggregate: sum \xi_j within (firm, year) to build the firm-year innovation-value panel — the central data object. 4. Sanity checks: (i) print the top-ten firm-years by total value and ask whether the names make sense (do the famous innovators show up?); (ii) note which patents lack a value and why — patents of private/foreign/non-listed assignees or without usable CRSP returns around the grant date get no value, the §6 selection made visible.

(b) (4 pts) Rebuild and stress one patent’s value (§7 Exercise 2). Pipeline (the Problem 2 calculation, automated): 1. Pick one patent and its owning firm; pull the firm’s daily returns. 2. Fit the market model R_{ft}=a_f+b_fR_{mt}+\varepsilon on a pre-event estimation window (e.g., ~120 trading days) ending a short gap (e.g., ~10 days) before the grant date, so event-window returns never leak into the estimation of the normal-return model. 3. Form the daily abnormal returns over the short grant window and sum to the CAR. 4. Multiply CAR by market cap at grant to get dollars; (optionally) inflate by 1/\pi for grant-anticipation; this is your hand-built \xi_j. 5. Compare your \xi_j to the published KPSS value for that patent. Then vary on purpose the two construction choices §7 flags — the window length (1-day / 3-day / 5-day) and the normal-return model — and watch the estimate move, the Problem 2(d) lesson that the measure depends on modeling choices.

(c) (2 pts) Validate, then break it; and data discipline (§7 Exercise 3 + Conventions §5). The validation regression: regress a future outcome (next year’s profitability/growth, or the known “true” value in synthetic data) on this year’s firm-year KPSS value, and confirm the positive, significant slope. The “break it” manipulation: inject simulated earnings-announcement returns into a subset of grant windows, rebuild the measure on the contaminated windows, and watch the validation degrade (Problem 4(b)). Data discipline: the notebook must (i) ship a seeded synthetic fallback with the same schema (simulated grants, dates, daily returns, and a baked-in “true” value) so every step runs end-to-end on a student laptop with no licensed data; and (ii) be honest about not fabricating real numbers — real CRSP daily returns stay read-only on GMU infrastructure with a pinned snapshot date (Conventions §5), the public KPSS dataset is a documented Path-A pull, and no real empirical estimates are invented.

All numerical answers verified in Python. Key results: P2 daily AR=\{0.76,1.25,0.37\}\% (predicted piece \hat b_fR_{mt}=\{0.44,0.55,0.33\}\%); 3-day CAR =2.38\%, gross =0.0238\times\$8\text{B}=\$190{,}400{,}000; with \pi=0.5, \xi_j=\$380{,}800{,}000; window stress gross dollars \$100.0\text{M} (1-day, 1.25\%) / \$190.4\text{M} (3-day, 2.38\%) / \$178.8\text{M} (5-day, 2.235\%). P5 shrink factors 1/\sqrt n=\{1.00,0.50,0.20,0.10\} for n=\{1,4,25,100\}; firm-year SNR \sim\sqrt n. Citations by name only, per chapter: Kogan, Papanikolaou, Seru & Stoffman (2017). All dollar figures and validation t-stats illustrative/constructed; no fabricated KPSS table numbers.

# Solutions — PS 6.1 (Building the KPSS Patent-Value Measure: Construction, Validation, and Critique) **Problem set:** `book/weeks/week-06/ps6.1.md` (PS 6.1, Week 6). **Chapter:** Ch 6.1 — Reader's Guide: *Technological Innovation, Resource Allocation, and Growth* (Kogan, Papanikolaou, Seru & Stoffman 2017), read through the seven-part Reader's-Guide anatomy. These are full worked solutions, not just answers. Notation follows `CONVENTIONS.md` and locks to Ch 6.1 and the prior weeks it builds on: the **market model** $R_{it}=a_i+b_iR_{mt}+\varepsilon_{it}$ fit on a pre-event **estimation window**; the day-by-day **abnormal return** $AR_{ft_k}=R_{ft_k}-(\hat a_f+\hat b_f R_{mt_k})$; the **cumulative abnormal return** $\text{CAR}=\sum_k AR_{ft_k}$ over the short grant window; the dollar conversion by **market capitalization**; the §2 three-refinement construction (filter the market → scale by market cap → inflate by $1/\pi$ for grant-anticipation) giving the per-patent value $\xi_j$; the Week-1 **construct-vs-measure** gap, **convergent/predictive validity**, **selection**, and **external validity**; and the variance-of-an-average shrinkage $\sigma\mapsto\sigma/\sqrt n$. Every numerical result here was confirmed in Python (NumPy abnormal-return and CAR arithmetic, the three-window stress, and the $1/\sqrt n$ table); verifying notes appear where useful. All dollar figures and table values are **illustrative and constructed**, never quoted as KPSS's own; the paper is cited by name as **Kogan, Papanikolaou, Seru & Stoffman (2017)** (KPSS). --- ## Problem 1 — Why a dollar measure? Counts, citations, and the construct–measure gap (14 points) **(a) (4 pts)** The **construct** Leah cares about is **the private economic value of a single invention — how many dollars the patent is worth to the firm that owns it.** The two standard proxies fail in distinct ways: *Patent counts* weight every patent **equally** — one trivial design tweak counts the same as the patent that defined a decade — so a firm with a hundred throwaway patents looks more innovative than a firm with the one transformative one. Counts measure *quantity*, not *value*. *Forward citations* are a real improvement (they track importance), but they carry **two** defects §1 insists you recite. First, *timing/speed*: citations accumulate over ten or more years, so you cannot value a recent patent in real time. Second, *truncation*: recent patents look artificially uncited simply because the future that would cite them has not happened yet. And neither proxy is denominated in dollars — the unit anyone actually cares about. **(b) (4 pts)** The clever move (§5) is to **let the stock market value the patent for you.** (i) A patent grant is *news* because the USPTO's decision and exact claims were uncertain beforehand; the grant resolves some of that uncertainty, and the market updates the firm's price the moment the grant becomes public. (ii) A large positive reaction says the market judged the patent valuable; a near-zero reaction says it judged it worthless — the size of the jump *is* the market's dollar estimate. (iii) This **outsources the valuation** to the thousands of investors who collectively set the price and have already done the hard work, so KPSS need *not* build a structural model of demand, competition, and discount rates (assumptions nobody can defend); they just read the price reaction with the cleanest tool available, the abnormal return in a tight window. **(c) (3 pts)** The *units* advantage: citations are not denominated in **money** — they count attention, not dollars — whereas the market reaction is a dollar figure directly. The *speed/timing* advantage: a stock reaction happens in **three days** and is available the instant the patent issues, while citations take **a decade** to accrue (and are truncated for recent patents). The speed is what makes a **long firm-year panel possible**: because every patent — including last year's — gets a value *immediately*, KPSS can build a dense year-by-year panel, and a panel is exactly what you need to study growth and reallocation, which play out over time. Citations could never deliver a real-time panel. **(d) (3 pts)** In Week-1 language, the **construct** is the patent's true economic value (unobservable), and the **measure** is the grant-window abnormal stock return, scaled to dollars (observable). The honest one-sentence description: *the KPSS number is the **market's** **expected** private value of the patent **at grant**.* Dropping any of the three words overclaims: drop "market's" and you pretend it is the true value rather than the market's opinion of it; drop "expected" and you pretend it is realized value rather than a forecast; drop "at grant" and you pretend it incorporates everything learned later. It is **not** the value the patent ultimately realized — a patent the market loved at grant that later flopped still carries a high KPSS value, which is precisely the construct-vs-measure gap. --- ## Problem 2 — Rebuild one patent's value from a grant-window event study (24 points) Given: $\hat a_f=0.0000$, $\hat b_f=1.10$, $\text{MktCap}_f=\$8{,}000{,}000{,}000$. Predicted normal return each day is $\hat R_{ft}=\hat a_f+\hat b_f R_{mt}=1.10\,R_{mt}$. **(a) (6 pts)** Abnormal return $AR_{ft_k}=R_{ft_k}-(\hat a_f+\hat b_f R_{mt_k})$, day by day: | $k$ | $R_{ft}$ | $R_{mt}$ | $\hat b_f R_{mt}$ | $AR_{ft_k}=R_{ft}-\hat b_f R_{mt}$ | |:---:|:---:|:---:|:---:|:---:| | $-1$ | $0.0120$ | $0.0040$ | $1.10(0.0040)=0.0044$ | $0.0120-0.0044=0.0076$ ($0.76\%$) | | $0$ | $0.0180$ | $0.0050$ | $1.10(0.0050)=0.0055$ | $0.0180-0.0055=0.0125$ ($1.25\%$) | | $+1$ | $0.0070$ | $0.0030$ | $1.10(0.0030)=0.0033$ | $0.0070-0.0033=0.0037$ ($0.37\%$) | (With $\hat a_f=0$ the intercept drops out.) You subtract $\hat b_f R_{mt}$ because of the §2 first refinement, **"filter out the market"**: a stock can move during the window for reasons that have nothing to do with the patent — the whole market rallied, or the firm's sector moved — and that contribution lives in $\hat b_f R_{mt}$. If you skipped it, the *raw* window return would mix the patent's value with whatever the broad market did those three days; the patent's value lives in the residual (the abnormal return), not the raw return. **(b) (5 pts)** Cumulative abnormal return over the three window days: $$ \text{CAR}=\sum_{k=-1}^{+1}AR_{ft_k}=0.0076+0.0125+0.0037=0.0238=2.38\%. $$ Gross dollar reaction: $$ \text{CAR}\times\text{MktCap}_f=0.0238\times\$8{,}000{,}000{,}000=\$190{,}400{,}000. $$ The market-cap multiplication is §2's second refinement, **"scale by market cap."** A raw percentage cannot be compared across firms of different sizes because the *same* 2.38% jump is a very different dollar amount for an \$8B firm than for a \$200M firm; multiplying by market cap converts the firm-specific percentage surprise into the dollars the market added to the firm's value on news of the patent. **(c) (5 pts)** The third refinement inflates for grant-anticipation. With $\pi=0.5$: $$ \xi_j=\frac{\text{CAR}\times\text{MktCap}_f}{\pi}=\frac{\$190{,}400{,}000}{0.5}=\$380{,}800{,}000. $$ The adjustment **raises** the estimate because a grant is only *partly* a surprise — markets anticipate some grants, so the observed window reaction captures only the *unanticipated* fraction of the patent's value; dividing by $\pi<1$ grosses the reaction back up to the full expected value. Interpretation, the careful sentence now attached to a number: **$\xi_j\approx\$381$ million is the market's expected private value, at grant, of patent $j$ to firm $f$** — not what the patent ultimately earned. **(d) (5 pts) Window stress.** Holding $\hat a_f,\hat b_f$ fixed and converting *gross* (pre-$\pi$) reactions to dollars: - **One-day** ($k=0$ only): $\text{CAR}_1=AR_{f,0}=1.25\%=0.0125$, so $0.0125\times\$8\text{B}=\$100{,}000{,}000.$ - **Three-day** ($k\in\{-1,0,+1\}$): $\text{CAR}_3=2.38\%$, so $\$190{,}400{,}000$ (from part b). - **Five-day** ($k\in\{-2,\dots,+2\}$): add $AR_{f,-2}=+0.005\%=0.00005$ and $AR_{f,+2}=-0.15\%=-0.0015$ to the three-day CAR: $\text{CAR}_5=0.0238+0.00005-0.0015=0.02235=2.235\%$, so $0.02235\times\$8\text{B}=\$178{,}800{,}000.$ Side by side: **\$100M (1-day) · \$190.4M (3-day) · \$178.8M (5-day)** — the estimate nearly *doubles* between the one-day and three-day windows and then drifts back down when two more days are added. This makes §6's vulnerability **"the measure depends on modeling choices"** concrete: the same patent's "value" moves by tens of millions of dollars purely because of the window you pick. The **short** window is the principled default because the longer you make the window, the more time you give *other* value-relevant news (earnings, deals, analyst revisions) to land inside it and contaminate the abnormal return — the very confound of §6 / Problem 4(b). So no single $\xi_j$ should be read as "the patent is worth exactly \$X"; it is a noisy, model-dependent point estimate, which is exactly why the firm-year aggregation of Problem 5 is the unit the paper actually trusts. *(Verification: $AR$ = $\{0.76,1.25,0.37\}\%$; CAR$_3$=2.38%, \$190.4M; CAR$_1$=1.25%, \$100.0M; CAR$_5$=2.235%, \$178.8M; $\xi_j$=\$380.8M at $\pi=0.5$. Confirmed in Python.)* --- ## Problem 3 — What would convince you the measure is valid? (18 points) **(a) (5 pts) Convergent validity.** This is §4 Step 1's first sanity test — **"does the market-based value correlate with forward citations?"** You wanted a **positive, significant** slope, and Leah's $+0.31$ ($t=5.8$) delivers it. The logic of why it is reassuring but not sufficient: citations are the literature's established proxy for patent importance, trustworthy *where they apply*, so if the dollar value rises with citations the new measure **agrees with the old one where the old one is trustworthy** — that is convergent validity, evidence the new number is tracking something real and not noise. But it is not sufficient because a measure that merely *reproduced* citations would add nothing — citations already exist, and they are slow and unit-less; the point of KPSS is to do *more* than citations (dollars, in real time), so convergent validity is the floor, not the ceiling. (Acceptable also: citations themselves are an imperfect proxy, so agreement with them cannot by itself certify agreement with the *construct*.) **(b) (5 pts) Predictive validity.** §4 calls the profitability/growth test "the strongest" because it shows the measure — built from a **three-day stock reaction** — forecasts a **real outcome years out**, the owning firm's future profits/output. In Week-1 terms: the citation test in (a) shows the measure **agrees with another proxy** (one imperfect ruler matching another), whereas the profitability test shows the measure **agrees with the construct itself** — economic value should, if real, be followed by higher realized profits, so a positive predictive slope is direct evidence that the number captures value, not just attention. Leah's $t\approx3.6$ (predicting a planted next-year outcome in `nb6.1`) is exactly this stronger evidence: a forecast of reality, not a correlation with a sibling proxy. **(c) (4 pts) The discriminant check.** The §4 test is **"does it survive the obvious confound?"** — showing the patent-grant signal is *distinct* from a generic announcement return rather than just reflecting *other* contemporaneous news. One concrete check on synthetic data: Leah compares the average grant-window abnormal return to the average abnormal return over **matched non-grant windows** for the same firms (placebo windows on ordinary days with no patent news) — if grant windows show a systematically larger reaction than placebo windows, the signal is attributable to the patent rather than to generic announcement-day churn. (Acceptable also: compare across patents that *do* vs *do not* later prove valuable, or regress out earnings-announcement returns and check the grant effect survives.) **(d) (4 pts) Falsifiability.** §4's rule for reading these tables is to read them for **signs and significance**, **not** for magnitudes you memorize — the pattern that matters is that the value measure is positively and significantly related to citations, profits, and growth. The measure would **break** ("the paper collapses here") if that pattern *failed*: if the value–citation slope were flat or wrong-signed, if the value–profit forecast were insignificant, or if grant-window returns were indistinguishable from generic announcement returns — any of these would mean the constructed dollar value is not tracking patent value at all. This entire validation exercise lives in Reader's-Guide section **§4, "Table-by-table reading order"** (specifically Step 1, the measure-validation evidence a referee reads *first*). --- ## Problem 4 — Read like a referee: three named vulnerabilities (24 points) **(a) (8 pts) Expectations, not realized value (Week 1 — measurement validity).** (i) The grant-window return is a forecast **of the patent's economic value**, formed by the market in three days using only the information available *then*. It is **not** the value the patent *ultimately realized*. The gap is exactly the Week-1 construct-vs-measure gap: the **construct** is true realized economic value; the **measure** is a three-day price reaction, i.e., the market's *expectation* of that value at grant. (ii) Concrete example, Leah's terms: a glamorous AI-chip patent grant that the market greets with a big positive abnormal return — earning a high $\xi_j$ — but whose technology is leapfrogged within two years and never earns the firm a dollar. For "flashy" patents the market over-loves at grant, the measure's error is **upward**: $\xi_j$ is too high relative to realized value (and symmetrically, quiet patents the market underrates carry a downward-biased $\xi_j$). (iii) The check (from the "Read like a referee" box): regress a *later realized* outcome — eventual profits attributable to the patent, eventual forward citations, or litigation/renewal outcomes — on the grant-window $\xi_j$, and ask whether the forecast is **unbiased** (slope $\approx1$ on average, errors centered at zero) or **systematically too high for some kinds of patents** (e.g., a negative residual concentrated among the flashiest, most-hyped technology classes). The deep reason the measure can **never, by itself**, separate "the market was right" from "the market was wrong but consistently wrong": the measure *is* the market's forecast, so any test using *only* the measure is circular — you need an *external* realized-outcome benchmark (the later data above) to grade the forecast, and even then a *consistent* bias would just shift the whole calibration without revealing itself within the grant-window number alone. **(b) (8 pts) Confounding news in the window (Week 4 — event-study identification).** (i) The identifying assumption (§2, one sentence): **the abnormal return in the grant window equals the market's unbiased estimate of the patent's value — as long as the patent grant is the *only* value-relevant news in that short window** (and markets are efficient enough to impound it). (ii) Two concrete other-news kinds: an **earnings announcement** (or a guidance update) landing in the window, and a **merger/acquisition or major-deal announcement** (also acceptable: an analyst rating revision, a lawsuit outcome, a regulatory ruling). Either corrupts $\xi_j$ because the window's abnormal return now mixes the patent's value with the *other* news's value — the price moved for two reasons and the construction credits all of it to the patent, biasing $\xi_j$ up or down by the size of the contaminating surprise. (iii) **Idiosyncratic** confounds are news items that hit individual patents in *random*, *uncorrelated* directions — some windows catch good news, some bad, with no relationship to the patent's true value. Averaging over the many patents in a firm-year (Problem 5) makes these cancel, because their mean tends to zero. **Systematic** confounds are contamination that is *correlated* across patents or *correlated with the true value itself* — e.g., if firms strategically time good-news disclosures near grant dates, or if grant windows for genuinely valuable patents also tend to carry positive earnings surprises. These do **not** wash out under aggregation, because their mean does *not* go to zero — averaging many same-signed errors leaves the bias intact. The systematic case is the genuinely dangerous one, and it is exactly the `nb6.1` "break it" extension where injecting contamination *correlated with* the true value survives aggregation and degrades the validation (whereas purely *random* contamination is averaged away even at high contamination shares). **(c) (8 pts) Selection and efficient markets (Week 1 — external validity; circularity).** (i) The three stacked selection screens (§6): **(1) only *granted* patents** get a value — rejected/abandoned applications, and the decision to even apply, are invisible; **(2) only patents of *publicly traded* firms with usable CRSP returns** can be valued — private firms, pre-IPO startups, universities, and foreign assignees drop out; **(3) the firm must *survive* and stay listed** around the grant date. The population the firm-year panel actually represents: **the innovation of large, public, surviving U.S. firms** — *not* "all innovation in the economy." (ii) A conclusion not extendable to the whole economy: that **a surge of innovation reallocates resources away from rivals across the economy** (the creative-destruction story at the aggregate level). The panel only sees public, surviving firms, so it is silent on innovation by private firms, startups, and universities — exactly the actors who do much of the disruptive innovating — so an economy-wide reallocation claim requires an argument the panel's selected sample cannot supply. (iii) The **circularity**: the measure is *built* on the assumption that the market prices the patent correctly and promptly at grant — efficiency is the *premise* that turns the abnormal return into "the value." So you cannot turn around and use the measure to *test* market efficiency, because you would be assuming the conclusion. "Efficient markets cuts both ways" for the measure's accuracy: if markets are **slow/noisy**, value leaks in *after* the window and the measure is **attenuated** (too low); if markets **over-react**, the window over-counts and the measure is **inflated** (too high). Either way the measure inherits whatever pricing errors the market makes, and the construction has no internal way to detect them. --- ## Problem 5 — Aggregation: why summing noise sharpens the signal (10 points) **(a) (5 pts)** Write $\hat\xi_j=\xi_j^{\text{true}}+e_j$ with $e_j$ roughly independent across patents and $\operatorname{Var}(e_j)=\sigma^2$. For the $n$ patents a firm is granted in a year, the **average** noise is $\bar e=\frac1n\sum_{j=1}^n e_j$. By the variance-of-an-average rule, with independent $e_j$, $$ \operatorname{Var}(\bar e)=\frac{1}{n^2}\sum_{j=1}^n\operatorname{Var}(e_j)=\frac{n\sigma^2}{n^2}=\frac{\sigma^2}{n} \;\Longrightarrow\; \operatorname{sd}(\bar e)=\frac{\sigma}{\sqrt n}. $$ So averaging $n$ roughly-independent noisy values shrinks the noise standard deviation by the factor $1/\sqrt n$: | Patents in firm-year, $n$ | $1$ | $4$ | $25$ | $100$ | |:---:|:---:|:---:|:---:|:---:| | Noise-SD shrink factor $1/\sqrt{n}$ | $1.00$ | $0.50$ | $0.20$ | $0.10$ | (A firm with 100 granted patents averages its per-patent noise down to one-tenth of a single patent's noise SD. Equivalently, the firm-year *total* has signal-to-noise growing like $\sqrt n$: the true total scales with $n$ while its noise SD scales only with $\sqrt n\,\sigma$, so SNR $\sim n/(\sqrt n\,\sigma)=\sqrt n/\sigma$.) **(b) (3 pts)** The shrinkage requires the noise $e_j$ to be **roughly independent (uncorrelated) across patents with mean zero** — only then do the cross terms in the variance vanish and the errors cancel rather than reinforce. This is the *same* distinction as Problem 4(b): aggregation defeats **idiosyncratic** confounds (independent, mean-zero window noise — which is what makes $\bar e\to0$), and it leaves **systematic** confounds untouched (correlated or value-dependent contamination, whose mean does *not* go to zero, so $\bar e$ stays biased). It is the same fork because whether the firm-year panel can be trusted comes down to exactly whether the per-patent errors are independent (washed out by summing) or systematic (carried straight through the sum). **(c) (2 pts)** Leah's firm-year correlation ($\approx0.72$) beats her per-patent correlation ($\approx0.4$) because the construction *is* unchanged but the **unit** changed: at the patent level each $\hat\xi_j$ carries the full per-patent noise SD $\sigma$ (window confounds, model error swamping a small true blip), whereas summing within firm-year averages that noise down by $1/\sqrt n$, so the firm-year totals track the true firm-year value far more tightly. The design lesson for her capstone: when a per-unit measure is **noisy but unbiased**, aggregating to a coarser unit buys you a **much cleaner signal** (noise falls like $1/\sqrt n$ while the true signal accumulates), at the **cost of resolution** — the firm-year panel can no longer answer *patent-level* questions (which *specific* patent was the blockbuster), only firm-year ones. --- ## Problem 6 — Design the replication pipeline for `nb6.1` (10 points) **(a) (4 pts) Inspect and aggregate (§7 Exercise 1).** Ordered steps: 1. Load the patent-level value file (one row per granted patent: firm id, grant year, $\xi_j$), or the seeded synthetic stand-in with the same schema. 2. **Plot the distribution of per-patent values first**, *before* trusting them — expect it to be **wildly right-skewed**, a few blockbuster patents dwarfing thousands of small ones (in `nb6.1`'s synthetic data the top decile holds the large majority of total value). That skew *is* the whole motivation for not counting patents equally: if value were roughly uniform across patents, counts would suffice; because it is heavy-tailed, one transformative patent outweighs a hundred trivial ones, so you must *weight by dollars*. 3. **Aggregate**: sum $\xi_j$ within (firm, year) to build the firm-year innovation-value panel — the central data object. 4. Sanity checks: (i) print the **top-ten firm-years by total value** and ask whether the names make sense (do the famous innovators show up?); (ii) note **which patents lack a value and why** — patents of private/foreign/non-listed assignees or without usable CRSP returns around the grant date get *no* value, the §6 selection made visible. **(b) (4 pts) Rebuild and stress one patent's value (§7 Exercise 2).** Pipeline (the Problem 2 calculation, automated): 1. Pick one patent and its owning firm; pull the firm's **daily** returns. 2. Fit the **market model** $R_{ft}=a_f+b_fR_{mt}+\varepsilon$ on a **pre-event estimation window** (e.g., ~120 trading days) ending a short **gap** (e.g., ~10 days) before the grant date, so event-window returns never leak into the estimation of the normal-return model. 3. Form the daily **abnormal returns** over the short **grant window** and sum to the **CAR**. 4. Multiply CAR by **market cap** at grant to get dollars; (optionally) inflate by $1/\pi$ for grant-anticipation; this is your hand-built $\xi_j$. 5. **Compare** your $\xi_j$ to the published KPSS value for that patent. Then **vary on purpose** the two construction choices §7 flags — the **window length** (1-day / 3-day / 5-day) and the **normal-return model** — and watch the estimate move, the Problem 2(d) lesson that the measure depends on modeling choices. **(c) (2 pts) Validate, then break it; and data discipline (§7 Exercise 3 + Conventions §5).** The validation regression: regress a **future outcome** (next year's profitability/growth, or the known "true" value in synthetic data) on **this year's** firm-year KPSS value, and confirm the positive, significant slope. The **"break it"** manipulation: inject simulated **earnings-announcement returns** into a subset of grant windows, rebuild the measure on the contaminated windows, and watch the validation degrade (Problem 4(b)). Data discipline: the notebook must (i) ship a **seeded synthetic fallback** with the same schema (simulated grants, dates, daily returns, and a baked-in "true" value) so every step runs end-to-end on a student laptop with **no licensed data**; and (ii) be honest about not fabricating real numbers — real CRSP daily returns stay **read-only on GMU infrastructure** with a **pinned snapshot date** (Conventions §5), the public KPSS dataset is a documented **Path-A pull**, and no real empirical estimates are invented. --- *All numerical answers verified in Python. Key results: P2 daily $AR=\{0.76,1.25,0.37\}\%$ (predicted piece $\hat b_fR_{mt}=\{0.44,0.55,0.33\}\%$); 3-day CAR $=2.38\%$, gross $=0.0238\times\$8\text{B}=\$190{,}400{,}000$; with $\pi=0.5$, $\xi_j=\$380{,}800{,}000$; window stress gross dollars $\$100.0\text{M}$ (1-day, $1.25\%$) / $\$190.4\text{M}$ (3-day, $2.38\%$) / $\$178.8\text{M}$ (5-day, $2.235\%$). P5 shrink factors $1/\sqrt n=\{1.00,0.50,0.20,0.10\}$ for $n=\{1,4,25,100\}$; firm-year SNR $\sim\sqrt n$. Citations by name only, per chapter: Kogan, Papanikolaou, Seru & Stoffman (2017). All dollar figures and validation t-stats illustrative/constructed; no fabricated KPSS table numbers.*