The effect of going public on stock prices: long-run abnormal returns after IPOs

Headline estimate. U.S. common-stock IPOs from Jay Ritter's IPO database, 1975–2021, earn a mean 3-year buy-and-hold abnormal return (BHAR) of −19.7% relative to the CRSP value-weighted market (skewness-adjusted t = −7.2; bootstrap 95% CI [−23.2%, −16.1%]). The median IPO does far worse (−55.7%); 70.5% underperform the market. Equivalently, a dollar in the average IPO grows to $0.86 per dollar in the market over three years (wealth relative = 0.862). This is the "new-issues puzzle" (Ritter 1991; Loughran-Ritter 1995): going public is followed by significant long-run underperformance.

Full numbers: output/results_table.md.

CAR cross-check. A cumulative-abnormal-return (CAR, additive) analysis tells the same story: the mean 36-month CAAR is −17.8% vs the VW market (cross-sectional t = −14.9; 54.3% of IPOs negative). CAAR is small and positive for the first ~3 months (a brief post-IPO drift) then declines steadily. It is slightly less negative than BHAR because summing returns does not compound the typical IPO's losses (firm-level corr ≈ 0.61). See output/car_table.md and the figure below.

---

What this does

1. Builds the event sample from Ritter's IPO-age.xlsx (offer dates + CRSP permnos).
2. Validates Ritter's permnos against CRSP and keeps U.S. common stock (shrcd ∈ {10,11}).
3. Pulls CRSP monthly returns, delisting returns, and the market index (legacy SIZ tables).
4. Computes delisting-adjusted, event-time buy-and-hold abnormal returns over months 1–36.
5. Produces the BHAR results table and figure with skewness-adjusted and bootstrap inference.
6. Runs a companion cumulative-abnormal-return (CAR/CAAR) analysis with its own table/figure.
7. Validates the constructed measures against Ritter's own published series (IPOALL.xlsx): IPO counts and first-day initial returns (see "Validation against Ritter" below).
8. Reconstructs the analysis from CRSP daily data, aggregated up to monthly, so the IPO month is included and a higher-frequency view is possible (see "Daily reconstruction" below).
9. Recovers true first-day (initial) returns by bringing in IPO offer prices — from Ritter's 1975–84 IPO2609.xls (stage 8), Capital IQ (stage 9), and parsed SEC 424B prospectuses for 1996+ (stage 10). See "First-day returns via offer prices" below.
10. Unifies those three sources into one per-IPO first-day-return series with provenance (stage 11) → data/processed/firstday_returns.parquet.

Stages 3b/4b/5b (daily) run alongside the monthly stages 3/4/5.

Method choices (BHAR vs CAR, benchmark, windows, delisting, matching) are each argued in DECISIONS.md; a stage-by-stage narrative is in LOG.md; and a sample-attrition table is written for every stage (output/attrition_*.csv).

Reproduce

pip install -r requirements.txt      # or: make deps
./run.sh                             # or: make all

Outputs land in output/. Re-running is deterministic (bootstrap is seeded).

What a collaborator must supply

• Their own WRDS access. Create a .pgpass file (libpq format, one line: host:port:dbname:user:password, e.g. wrds-pgdata.wharton.upenn.edu:9737:wrds:USER:PW) and either place it at ./.pgpass or set PGPASSFILE=/path/to/.pgpass. The .pgpass file is git-ignored and must never be committed.
• The Ritter file is auto-downloaded from https://site.warrington.ufl.edu/ritter/ipo-data/ (IPO-age.xlsx); if the site is unreachable, download it manually into data/raw/.
• Raw WRDS extracts and the Ritter file are git-ignored (data/); only code, docs, and the final figure/table are tracked.

Environment

Python 3.11; pandas 2.x, numpy, matplotlib, scipy, psycopg2, openpyxl (see requirements.txt). CRSP monthly data on the WRDS instance used ran 1925-12 .. 2024-12, which sets the 2021 sample cutoff (every cohort gets a full 36-month window).

How the Ritter→CRSP match was done

Ritter's IPO-age file already contains a CRSP Perm (permno) for each IPO, so we match on permno — never tickers (which are recycled). We validate every permno against crsp.msenames and cross-check Ritter's CUSIP against CRSP ncusip. Match results:

step	N
Ritter rows (IPO-age)	16,030
with a CRSP permno, 1975–2021, non-ADR, unique	14,084
permno present in CRSP	14,080
U.S. common stock (shrcd ∈ {10,11})	11,283
after $5 penny screen + ≥1 aftermarket month	10,101

CUSIP is sparse in Ritter (so it is not the key), but conditional on Ritter carrying a CUSIP, 6-char issuer agreement with CRSP is 95.6%, confirming the permnos are correct.

Validation against Ritter's published measures (issue #1)

Jay Ritter publishes his own constructed IPO measures in IPOALL.xlsx (monthly IPO counts and average first-day returns). Stage 7 (src/s07_validate_ritter.py) checks ours against his; full numbers in output/ritter_validation_table.md.

• Event universe — validated. Our final sample of 10,101 IPOs (1975–2021) is bracketed by Ritter's restrictive net count (9,199) and gross count (15,905), and the year-by-year counts track his almost one-for-one: corr = 0.96 vs the net count (0.96 vs gross). We picked up the right events.
• First-day initial returns — not recoverable from CRSP monthly data, by construction. In the IPO month, 95.7% of our firms have a CRSP closing price but only 0.1% have a CRSP monthly return: CRSP does not compute the first partial-month return (no prior month-end base price, and the offer price is not used), so the offer→first-close "pop" that is Ritter's first-day return is absent from ret. Our nearest measure, the first full aftermarket month (event month 1), averages 1.6% vs Ritter's 17.3% first-day average (annual corr 0.38). This is precisely the "CRSP issue" anticipated in the request — and exactly why the long-run analysis excludes the IPO month and starts at month 1. The underpricing pop sits outside our window, so it neither inflates nor contaminates the −19.7% three-year BHAR headline.

Daily reconstruction (issue #3)

Because the monthly file misses the IPO month (above), stages 3b/4b/5b pull CRSP daily returns, aggregate them up to monthly, and rebuild the analysis. Full numbers in output/daily_results_table.md.

• Consistency check passes. Over the same window (months 1–36), the daily→monthly reconstruction gives mean BHAR −20.0% vs the monthly analysis's −19.7% — agreement to a fraction of a point confirms the daily pull and aggregation are sound.
• The IPO month is now included. Daily CRSP carries a return from the first trading day, so we can start the event clock at month 0. The fuller months 0–36 window gives mean BHAR −20.5% (the recovered IPO-month return averages +1.5%) — the new-issues puzzle is unchanged.
• Higher-frequency view. A daily event-time CAAR over the first ~3 months shows the early post-IPO drift at daily resolution (+2.6% by trading day 63).
• First-day underpricing still isn't recoverable — even from daily data. CRSP has no offer price, and the pop is mostly the overnight offer→open move; CRSP's first secondary- market return averages just 0.5% vs Ritter's 17.3% (annual corr 0.33). Matching Ritter's first-day return requires his offer-price-based data, not CRSP. This refines #1: the monthly file misses the IPO month entirely; daily recovers the month but not the first-day pop — either way the underpricing is outside the long-run window and doesn't drive the headline.

First-day returns via offer prices

Issues #1/#3 showed the one thing CRSP lacks is the offer price, so the offer→first-close underpricing pop is invisible to CRSP at any frequency. Stages 8–9 bring the offer price in.

Stage 8 — Ritter IPO2609.xls (1975–84). The only public Ritter file with per-firm offer prices. Splicing its offer price onto the CRSP first-day close we already pull recovers a true first-day return of 9.7% (median 2.7%) over 1,163 firms — matching Ritter's own 8.3% (corr 0.96) — versus the CRSP-only first-day return of 0.4% (corr with Ritter just 0.12). Ritter's first aftermarket price equals our CRSP first close (corr 0.995), so the splice is clean. Numbers: output/offer_price_table.md.

Stage 9 — Capital IQ (post-1984). Ritter posts no offer prices after 1984, so for the modern era we use Capital IQ's ready-made first-day return (eqtypuboffr1dayretpct), linked to our sample via permno→gvkey→CIQ companyid (≈2,200 firms, mostly 1996+). Its annual median first-day return tracks Ritter's IPOALL series (corr 0.50); the field has a clean median (~12%) but a junk tail, so we screen outliers and report the median and a winsorized mean, never the raw mean. CRSP-only first-day returns for the same firms are ~0% (corr 0.01 with CIQ) — the pop is still outside CRSP. Numbers: output/capitaliq_table.md.

Stage 10 — SEC 424B prospectuses (1996+). The authoritative source: the final Rule 424(b) prospectus carries the offer price on its cover. We map each IPO permno→gvkey→CIK (Compustat), pull the firm's filings from EDGAR, take the 424B nearest the offer date, and parse the offer price off the cover (filings cached under data/interim/edgar/). We parsed an offer price for 4,572 of 5,444 (84%) 1996–2021 IPOs with a CIK — 4,429 kept after a plausibility screen (81% overall; 92% for 2010–2021). Splicing the offer price onto the CRSP first-day close yields a true first-day return whose annual mean tracks Ritter's IPOALL series with corr 0.80 — far better than Capital IQ (0.50) or CRSP-only (~0). This is the cleanest path for 1996–present and the natural complement to Ritter's 1975–84 file. (The permno→CIK link uses the CCM link nearest the offer date, not one required valid at it — an IPO's Compustat link usually starts at its first fiscal year-end, so the strict version dropped ~half the firms.) Numbers: output/edgar_table.md.

Stage 11 — unified series. Stitches the three offer-price sources into one per-IPO first-day return, best-source-per-firm (Ritter 1975–84 → SEC 424B 1996+ → Capital IQ fallback), in data/processed/firstday_returns.parquet alongside the CRSP-only first-day return and the provenance of each value. Coverage: 6,122 of 10,231 analysis IPOs (60% overall, but 97% of 1975–84, 87% of 1996–2009, 97% of 2010–21 — the 60% is dragged down only by the 1985–95 gap). Independent cross-check: SEC 424B and Capital IQ agree at corr 0.70 on 3,808 overlapping firms. Numbers: output/unified_firstday_table.md.

Still missing: offer prices for 1985–1995 (before EDGAR's electronic coverage and after Ritter's IPO2609). The remaining options are licensed SDC New Issues (not on this WRDS instance) or hand-keyed sources.

What I'm least sure about (most likely to change the result)

1. Benchmark = market, not a size/B-M matched firm. IPOs are small, high-growth firms; a style-matched benchmark (Loughran-Ritter's preferred control) typically shows more underperformance than the market. Using the equal-weighted market already shrinks the estimate to −9.6%, so the benchmark is the single biggest lever. A Compustat-based size×book-to-market control firm is the natural next step.
2. Delisting-return handling. We incorporate dlret and impute −30% for performance delists missing it (Shumway 1997). This affects only 21 firms, and the realized-horizon variant gives nearly the same answer (−19.1%), so this is unlikely to flip the result — but it is the classic place survivorship bias hides.
3. BHAR vs CAR and skewness. BHAR is right-skewed; the mean is sensitive to a few extreme winners (winsorizing pushes it to −25.1%, the median to −55.7%). We report skewness-adjusted and bootstrap inference, but a calendar-time-portfolio (Fama-French alpha) cross-check would be the most decisive robustness test.

Closing summary

Across 10,101 U.S. common-stock IPOs from Ritter's database (1975–2021), the average firm underperforms the CRSP value-weighted market by 19.7% over the three years after going public (median −55.7%; wealth relative 0.86), a result that is statistically strong under skewness-adjusted and bootstrap inference and robust to the delisting treatment, the return horizon, winsorizing, and using cumulative rather than buy-and-hold abnormal returns. The three judgment calls a reviewer should check first: (1) the benchmark — market-adjusted rather than size/book-to-market matched, which is the choice most likely to move the number; (2) the delisting-return convention — incorporating dlret and imputing −30% for performance delists, the usual home of survivorship bias; and (3) the abnormal-return construction — fixed 36-month BHAR reinvesting delisting proceeds in the market, with the known right-skew of BHAR addressed via skewness-adjusted/bootstrap statistics.