An ordinal logistic regression analysis of 796 A-share IPOs (2022–2025), examining how issuance characteristics, firm fundamentals, and market conditions shape the ordered classification of first-day returns.
A well-documented empirical regularity across global capital markets is IPO underpricing—shares offered at prices systematically below their first-day closing price (Ljungqvist, 2007).
Rock (1986) proposed the Winner’s Curse model: underpricing compensates uninformed investors facing adverse selection. Beatty & Ritter (1986) showed that greater ex ante uncertainty leads to higher underpricing, with investment banks’ reputation as enforcement. Ritter & Welch (2002) emphasized that IPO phenomena are non-stationary.
China’s A-share market historically exhibited average first-day returns exceeding 100% during certain periods (Chen, 2022), attributed to the approval-based system’s implicit 23× P/E ceiling—creating systematic mispricing and fueling speculation.
The CSRC launched registration-based reforms: STAR Market (2019) → ChiNext (2020) → full implementation across all boards (Feb 2023), removing administrative price controls, strengthening disclosure requirements, and shifting pricing power to the market.
Deng et al. find efficient STAR pricing, yet Li & Li show reform initially amplified speculation. Lyu documents a temporary efficiency dip before recovery. The relationship between reform and first-day performance is complex and evolving.
STAR Market registration regime offers the most efficient pricing with lower investor overreaction.
• Decomposed initial returns into fair value & overreaction components
Market-oriented reform initially magnified initial returns due to amplified speculation.
• DID approach; pricing efficiency improvements require time
Pricing efficiency declined in Year 1 of ChiNext reform, then gradually recovered.
• SFA model on 354 ChiNext & 119 STAR companies
Third-party government certification reduces information asymmetry under registration system.
• 1,312 firms (2019–2023); “Little Giant” effect most pronounced
Most studies treat IPO first-day returns as a continuous variable (OLS) or binary outcome (underpriced vs. overpriced). In practice, a stock that breaks issue price, one that doubles, and one that triples carry qualitatively different implications for investors.
Few studies have examined the full registration-era window (2022–2025) encompassing all board types under the unified system, including the 2023 full-rollout year.
Limited attention to how issuing characteristics interact with market-level conditions (e.g., does a high P/E issue perform differently in bull vs. bear markets?).
4 ordered categories preserving natural ordering: Break / Mild Rise / Significant Rise / Surge. Unlike binary or continuous approaches, this captures the qualitative gradient of investor outcomes.
796 non-financial IPOs across Main Board, ChiNext & STAR Market, Jan 2022–Dec 2025. Data sourced from Tushare Pro & cninfo/akshare, covering issuance records, first-day trading, financials, and market indices.
14 predictors spanning Issuance Firm Market dimensions, plus year & 33 industry fixed effects. Interactions: Board×PE, MktRet×PE.
972 initial records → −3 (no listing date) → −4 (financial sector) → −168 (BSE, no trading data) → −1 (out of period) → 796 Final Sample
13.9%
54.6%
23.5%
7.9%
| Category | Variable | Definition |
|---|---|---|
| Issuance | PE_issue | Issue P/E ratio |
| Price | Issue price (RMB) | |
| log(Funds) | ln(total proceeds) | |
| Ballot | Online lottery rate (%) | |
| Board | Listing board dummies | |
| Firm | Size | ln(total assets) |
| Lev | Debt-to-asset ratio (%) | |
| ROE | Return on equity (%) | |
| RevGrowth | Revenue growth (%) | |
| Market | MktRet_pre | CSI 300 return (20d) |
| Turnover | First-day turnover (%) | |
| UW_rep | Top-10 underwriter | |
| IPO_count | Concurrent IPOs (±30d) |
+ Year FE & 33 Industry FE in extended models. Continuous variables Winsorized at 1st/99th pctl.
Issuance characteristics only:
PE_issue, Price, log(Funds), Ballot, Board.
Tests whether offering structure alone predicts first-day category.
+ Firm fundamentals (Size, Lev, ROE, RevGrowth) + Market conditions (MktRet_pre, Turnover, UW_rep, IPO_count) + Year & Industry FE.
Core specification for hypothesis testing.
+ Board × PE_issue (does P/E effect differ by board?)
+ MktRet_pre × PE_issue (does market mood moderate P/E effect?)
Exploratory cross-level interactions.
The proportional odds assumption requires that coefficients β remain constant across all J−1 cumulative logits. Assessed via the Brant test (overall + per-variable). If rejected → re-estimate with Multinomial Logistic Regression as robustness check; consistency between the two lends confidence to the ordinal conclusions.
VIF: All < 3 (max = 2.90) — no multicollinearity.
Min category n: 63 observations (Surge) — well above the conventional ≥10 threshold.
Winsorization: Continuous variables trimmed at 1st & 99th percentiles.
Missing values: <17% per variable, imputed via industry-median.
Software: R polr() (MASS),
brant(),
multinom() (nnet).
Takeaway: Pricing efficiency improved dramatically—breaks vanished while surges emerged, reflecting full market adjustment to the registration system.
SD: 47.13
SD: 28.03
SD: 12.37
SD: 22.90
SD: 4.99%
Larger offerings constrain speculative upside, converging pricing toward intrinsic value. Coef: −2.011, p<0.01
Capital gravitates toward larger enterprises offering greater certainty under the registration regime. Coef: 0.799, p<0.01
Bullish sentiment dramatically amplifies first-day returns. Strongest single predictor in the model. Coef: 8.486, p<0.01
Higher first-day liquidity is significantly associated with better performance categories. Coef: 0.072, p<0.01
Oversubscribed IPOs with lower lottery rates show higher first-day returns. Coef: −5.921, p<0.01
Overall: χ² = 142.12, df = 36, p < 0.001
Key violations: Board dummies (ChiNext: χ²=83.23), Turnover (χ²=16.33),
Ballot (χ²=9.12), STAR (χ²=75.97), BSE (χ²=22.87).
Implication: The factors driving a “Break”
outcome are asymmetric to those propelling a “Surge”—the proportional odds
model may over-simplify the extremes.
Key findings confirmed:
• log(Funds): coef −1.052 to −3.154 across categories
• MktRet_pre: increasing effect (6.920 → 15.511)
• Turnover: progressive coefficients (0.319 → 0.609)
• 3-category & 5-category schemes yield consistent results
Re-estimated with alternative DV thresholds:
• 3-category: Merge Mild & Significant Rise → consistent signs & significance
• 5-category: Split Surge into two tiers → core conclusions hold
Results are not driven by arbitrary categorization.
Higher-performing categories exhibit systematically higher turnover, supporting the liquidity channel hypothesis. The ordinal model provides a useful summary, but the Brant test violation confirms asymmetric drivers across categories—multinomial results and sensitivity analyses both validate the core conclusions.
Our results corroborate the progressive improvement in pricing efficiency under the registration reform, aligning with Ritter & Welch’s (2002) assertion on the non-stationarity of IPO phenomena.
Fundraising scale (OR=0.134) constrains speculative upside. Ballot rate (OR=0.003) confirms oversubscription as a barometer of returns. Notably, the P/E ratio’s economic magnitude is trivial (OR=1.006)—after removing the 23× cap, the valuation multiple is no longer a core determinant.
Connects to: Lin & Zhang (2024)—market-oriented pricing is the most direct pathway to reducing underpricing.
Firm size (OR=2.224) shows a robust positive effect—counter-intuitive, as smaller-caps might be expected to surge more. Under registration, capital gravitates toward larger enterprises. ROE and leverage are insignificant: short-term trading is decoupled from profitability fundamentals.
Connects to: Qian et al. (2025)—financial indicators matter more for pricing efficiency than pricing-related factors.
STAR, ChiNext & BSE all negative vs. Main Board (ORs: 0.119–0.211), consistent with Lyu’s (2022) investor learning hypothesis. Market sentiment is the paramount amplifier (MktRet_pre OR=4,844). Break rate: 22.3% (2022) → 0% (2025) validates sustained pricing efficiency.
Connects to: Lyu (2022)—investor learning & rational calibration follow initial reform volatility.
Brant test rejects proportional odds (p<0.001): asymmetric drivers between “Break” and “Surge.” Multinomial logistic confirms turnover & sentiment effects are magnified in extreme right-tail categories—liquidity-driven frenzies.
Implication: The ordinal model provides a useful summary, but category-specific effects warrant further investigation.
Exclusion of 168 BSE IPOs due to data unavailability introduces selection bias. First-day price limits on the Main Board may also truncate observed returns. Future: Extend to full board spectrum once BSE data becomes available.
No micro-level book-building variables (institutional bid dispersion, pricing adjustments). DV categorization thresholds carry inherent subjectivity. Future: Integrate granular bidding data alongside underwriter adjustments.
Potential endogeneity between turnover and first-day returns not fully addressed. Behavioral underpinnings of extreme surges remain underexplored. Future: Social media sentiment, textual analysis, investor attention metrics.
The near elimination of price breaks signals the reform’s success. However, the 30.2% incidence of surges in 2025 underscores the need to monitor extreme positive deviations.
• Continue strengthening information disclosure
• Monitor IPO clustering effects on market saturation
• Consider circuit-breaker mechanisms for extreme first-day moves
The importance of fundraising scale and market timing has eclipsed simplistic “subscribe-to-all” strategies.
• Issuers: Smaller offerings + favorable market windows = better first-day reception
• Investors: Monitor pre-listing CSI 300 trends and IPO clustering density
• Underwriters: Reputation alone (UW_rep) no longer significantly predicts outcomes
This study adopts an ordinal categorical lens to capture the economic thresholds of IPO first-day performance, affirming that the registration reform has not merely deregulated pricing but has fundamentally reconfigured the determinants of initial returns.
SDSC 8013 — Statistical Methods for
Categorical Data Analysis
Semester B 2025/26
796 IPOs • 4 Years • 14 Predictors • 3 Models