The IPO isn’t
the finish line.
What ten years of tech IPOs can—and can’t—tell us about Anthropic.
Explore the evidenceWhat happens after the opening-day excitement?
We started with a question: how might an Anthropic IPO behave in its first 90 days, using the largest technology IPOs as historical comparisons?
That question grew into a broader study of U.S.-listed technology and digital businesses, their post-IPO price returns, and their performance against SPY and QQQ over the same dates. We extended the holding periods to six months and one year, then looked at which early winners stayed ahead.
The useful result is a framework for examining an IPO decision: the entry price, opportunity cost, range of outcomes, and quality of the evidence.
The typical IPO and the average IPO tell different stories.
At 90 days, the available-history sample had a +1.4% mean price return, while its median was −9.8%. A smaller number of large gains lifted the average. Only 35.7% beat SPY and 34.4% beat QQQ.
| Holding period | Histories | Beat SPY | Beat QQQ | Median IPO return |
|---|---|---|---|---|
| 90 days | 224 | 35.7% | 34.4% | −9.8% |
| 6 months | 215 | 34.4% | 34.0% | −12.6% |
| 1 year | 203 | 31.5% | 30.5% | −22.2% |
Different horizons contain different samples. Returns exclude dividends, trading costs and taxes. “Beat” means a strictly higher price return over paired dates.
Early winners didn’t always stay ahead.
Among the same 203 issuers with complete comparisons at all three horizons, 74 beat both benchmarks at 90 days. Of those, 39 still beat both at one year; 37 beat both at all three checkpoints. The 39 are selected retrospectively. They are useful to study, but are not a list we could have known in advance.
Longer holding alone did not produce broad outperformance.
In that fixed 203-issuer sample, the proportion beating SPY fell from 38.9% at 90 days to 35.0% at six months and 31.5% at one year. That comparison controls for the changing sample across horizons, but does not remove selection bias from missing histories.
Change the lens.
Select a benchmark and holding period, then inspect the observed sample. “Same issuers” keeps both benchmarks and all three horizons available for every included issuer.
Look inside the sample
| Company | IPO date | IPO return | Benchmark | Excess (pp) | Matched exit |
|---|
Search narrows the company table only. Aggregate metrics and chart describe the selected cohort and sample. Excess returns are percentage-point differences, not risk-adjusted alpha.
The frozen overview is shown above. The interactive explorer loads automatically when JavaScript is available.
A famous company can still be a difficult entry.
This study does not establish a reliable “buy after day X” rule. Every holding period starts at the first trading-day close. Testing an entry after 30, 60 or 90 days would require a separate analysis using information available at that later entry.
Before applying these comparisons to Anthropic, we would need its actual offering valuation, financial disclosures, share structure and float, lockup terms, and the price at which a public-market buyer could enter. The historical sample cannot supply those inputs.
A useful next research step is to test a specific, predeclared rule against suitable comparisons and an untouched holdout, including execution costs and realistic option availability. The model appendix below shows why keeping that boundary visible matters.
Follow the evidence.
Universe and coverage
The rolling census covers offerings dated October 26, 2016 through October 26, 2026. Its broad technology definition includes software, computing hardware, digital platforms, financial and education technology, and qualifying digital businesses. SPACs, funds, direct listings, and already-public cross-listings are excluded.
Across reconciled sources, 3,644 candidate offerings were classified: 588 inclusions and 3,056 exclusions. This is a source-reconciled research census, not an exchange-certified complete security master. Classification completeness is separate from price-history coverage.
Entry, endpoints and benchmarks
Entry is the first trading-day closing price, not the offer or opening price. Ninety days means calendar days; six and twelve months use calendar anniversaries. Each IPO/benchmark pair uses identical observed entry and exit dates. Exit closes are no later than the target anniversary and at most four calendar days stale; unacceptable histories remain missing.
Price returns exclude dividends, costs and taxes. Split and issuer-identity checks are upstream requirements. SPY and QQQ are ETF price-return comparators; excess return is a percentage-point difference. They do not establish causal or risk-adjusted performance.
Missing histories and retrospective selection
The extended study starts from the accepted 224-history 90-day pool. There are 215 six-month and 203 one-year matched histories. Some issuers are too recent for longer horizons; unavailable or unverifiable histories also restrict coverage. The constant sample requires both benchmarks at all three checkpoints.
Missing delisted histories can bias the results materially. The winner cohorts are selected using subsequent outcomes and cannot validate a prospective strategy. November subsets are small and cannot establish a seasonal edge.
Source ledger
Offering identities and classifications were reconciled from Jay Ritter’s IPO research, Nasdaq IPO calendars, IPOScoop issuer profiles, IPOBoutique pricing tables, Stock Analysis IPO lists, and issuer/SEC evidence where needed. Daily price histories were acquired from Yahoo Finance chart responses and checked for coverage and identity.
This publication contains derived checkpoint returns, aggregate statistics and research commentary. It does not distribute the underlying daily price archives, source workbooks, or a licensed historical options dataset. Source data remain subject to their providers’ terms. No third-party redistribution license is asserted.
Research snapshot and reproducibility
Version ipo-2026-10-26-v1 freezes the October 26 research snapshot, using prices through October 23. The public derived-results snapshot contains the source Git revision, SHA-256 hashes, paired checkpoint dates and returns, coverage counts, and model evidence. It contains no daily price series.
The publisher recomputes aggregate counts, wins, means and medians from the observations and rejects differences from the archived research summaries. Hashes establish snapshot consistency, not independent certification of vendor data. The source Git revision is an internal identifier, not a public repository link. The public snapshot supports independent recalculation of the derived summaries; reproducing the original acquisition requires obtaining the underlying data from its sources.
The allocation model did not pass its holdout checks.
The original conditional model compared shares, one cash-secured put tranche, and cash over 90 days. Its largest-ten historical cohort is separate from the ten-year census above: nine of ten primary issuers have accepted paths, with Agere missing.
| Evaluation | Mean 90-day return |
|---|---|
| In-sample selection | +2.4% |
| Leave-one-out selection procedure | −0.7% |
| Chronological held-out selection | −1.2% |
| Cash comparator (assumed 4% annual yield) | +1.0% |
Holdouts refit the selection procedure on training paths and evaluate the resulting allocations on held-out paths. They do not test the exact final in-sample allocation. The chronological holdout also breached the configured tail-loss constraint.
Illustrative inputs and interpretation
The initial model uses a $100,000 illustrative budget, $100 reference price, 80% assumed annualized implied volatility, 4% annual cash yield, 50% ownership cap, and 15% empirical tail-loss limit. It assumes one put tranche entered on day five with 30 days to expiry; commissions are $0.65 per contract and share-purchase slippage is five basis points.
Synthetic premiums use a haircut to theoretical option value, rather than observed American-option prices, volatility surfaces or fills. The empirical loss constraint is not a future loss guarantee. Nine historical scenarios cannot support a calibrated Anthropic forecast. These inputs are research assumptions, not the author’s account balances or personal allocation instructions.
The result worth sharing is the failed validation, alongside the model’s assumptions. It helps prevent an attractive in-sample result from becoming an unsupported trading claim.
Keep the question.
Keep the evidence with it.
Infinitrade is building a command center for portfolio context, research, market data and optional strategy operation—connected to the AI tools you already use.
Explore Infinitrade