Does Maturity Mismatch Explain When Large Firms Stop Being Safe?

Investors usually treat big companies as the safe part of the stock market. When a crisis hits, money moves out of small stocks and into large ones, a pattern called "flight to quality," and the extra return that small stocks normally earn disappears. There is a problem with this story. A large panel of Chinese listed firms shows that bigger firms are more likely to fund long-lived investments with short-term loans that must keep being renewed. This habit is called short-term financing for long-term investment (SFLI). It is a timing gap between cash owed and cash coming in, and it is the same kind of gap that sank Silicon Valley Bank. This proposal asks whether that hidden refinancing exposure decides which large firms actually receive the crisis-time safe-haven premium and which ones crash alongside small caps.

The core idea fits in two sentences. Size may not be the protective trait in itself. Size may act as a protective buffer only for firms that do not depend on rolling over short-term debt. If so, the familiar collapse of the small-stock premium in crises should come almost entirely from low-mismatch large firms. High-mismatch large firms should behave like small firms, but only during funding shocks, when credit gets scarce, and not during pure news shocks. That difference between two kinds of shock is what makes the claim testable and separates it from a generic "add another risk factor" exercise.

Core Claim

We claim that the size effect in stock returns under stress is conditional on balance-sheet maturity mismatch, and that the effect depends on the channel of the shock. There are three concrete predictions. First, in portfolios sorted jointly on market capitalization and SFLI, crisis-period outperformance of large caps will be concentrated in the low-SFLI large-cap cell. Second, the high-SFLI large-cap cell will underperform during episodes of tight short-term funding but not during information shocks such as trade-war news. Third, adding SFLI to a crash-warning model will raise accuracy for large caps, where current models are weakest, but not for small caps.

Several papers motivate the claim. The Impact of Digital Transformation on Corporate Short Term Lending and Long Term Investment measures SFLI from cash-flow statements for 11,292 A-share firm-years (2011–2021). It finds a significant positive association with firm size (standardized coefficient 0.017) and with leverage (0.010). It treats SFLI purely as a corporate-finance outcome and never looks at stock prices. We reuse its measure and ask how the market prices it. Several crisis studies document the flight-to-quality side:

Both studies treat "large" as a single group. Our proposal splits that group.

Two single-firm case studies show that scale does not protect against a timing gap. Research on the Risk Management of Pharmaceutical Circulation Enterprises shows Sinopharm carrying heavy short-term debt despite 311 billion yuan in assets, because receivables rose from 109 to 134 billion yuan. Silicon Valley Bank Bankruptcy traces a 6.2-year securities duration funded by largely uninsured deposits. Neither case study generalizes to the cross-section of stocks.

Two results constrain the design. Stock Price Crash Warning in the Chinese Security Market Using a Machine Learning-Based Method and Financial Indicators finds that large-cap crashes are driven mainly by long-term debt to assets and working capital over assets, while small-cap crashes are driven by asset growth. Any SFLI effect must therefore survive controls for those ratios. E2EAI finds that trading liquidity (turnover, spreads) matters little for large caps. That makes trading liquidity an unlikely explanation for large-cap stress behavior. It also has to be ruled out as an alternative explanation.

Research Program

The work has three parts, all on Chinese A-shares, where SFLI has already been measured.

Measurement. We rebuild the SFLI measure of Xu et al. annually from CSMAR cash-flow statements for 2011–2022. It captures long-term investment spending not covered by long-term financing. We lag it by one reporting cycle so that portfolios use only information already published. We assign each firm to the size tiers of Deng et al.: below CNY 5 billion, CNY 5–10 billion, and above CNY 10 billion.

Pricing. Each month we form independent 3×3 portfolios on size tier and SFLI tercile. We estimate Fama-French five-factor alphas and factor loadings for each cell. The five factors are market, size, value, profitability and investment. We then split the sample into two kinds of stress episode:

  • Funding-stress months, the channel the hypothesis says should matter. We identify these from spikes in short-term interbank rates, a series we would collect separately; it is not in the research file. The 2015 market crash and the 2020 pandemic onset are candidate episodes.
  • Information-shock days, the placebo channel. We take these from the trade-friction news calendar used in China-U.S. Trade Frictions, Opinion Divergence, and Stock Volatilities, which found that trade news moves large-cap volatility through investor disagreement.

The test statistic is the difference in abnormal returns between high-SFLI and low-SFLI large caps. We compare this difference across the two episode types.

Tail risk. We replicate the crash-label setup of Deng et al. on weekly data for 2015–2020 and retrain their XGBoost classifier for each tier, with and without SFLI. We use SHAP attributions to check whether SFLI takes over importance from the debt and working-capital ratios or adds information beyond them.

Baselines

Naive reference: unconditional size sorts. This is the standard three-factor view: sort on size alone and measure the crisis-period small-minus-large spread. It is the belief being challenged. We run it on the identical sample and episodes. The proposal beats it if the large-minus-small crisis spread becomes statistically indistinguishable from zero within the high-SFLI row while staying significant in the low-SFLI row. For the size factor's crisis-period magnitude in A-shares, Zhu reports no specific factor values, so no reported result was found on Lacuna for that quantity. As an indication of how poorly linear sorts behave locally, Research on Return Forecasting and Portfolio Construction in China's A-Share Market reports beta-sorted portfolio Sharpe ratios of −4.839 and −2.841.

Strongest existing method for tail risk: the XGBoost-NSGA-II-SHAP crash model of Deng et al. It is the only model in the file that already conditions crash prediction on size tier and balance-sheet ratios, so it is the fairest test of whether SFLI adds anything. The paper reports accuracy of 63.93% for large caps, 73.83% for medium caps and 78.41% for small caps. We rerun it ourselves on the same weeks, with the same feature set and the same multi-objective hyperparameter search budget. The only change is adding SFLI. Comparing against our own replication, not the published number, keeps data vintage fixed. Beating it means a large-cap accuracy gain larger than the variation across random seeds, with no comparable gain for small caps.

Strongest method for returns: E2EAI. It reports a CSI 1000 excess return of 20.12% and an information ratio of 4.35, against 16.36% for DMFM from Factor Investing with a Deep Multi-Factor Model. It is not trying to answer our question. It serves as a check on whether an SFLI group, added to its factor-masking inputs, receives nonzero weight for CSI 300 constituents during funding stress. We would train it with its published architecture and the same training windows.

Key ablation. We replace SFLI with the three variables most likely to proxy for it: leverage, working capital over assets, and investment growth, which is the investment factor's underlying characteristic. If the double-sort pattern and the crash-accuracy gain survive with these controls and vanish when SFLI is removed, the maturity-mismatch channel is isolated.

Why not just…?

Why not just use leverage and working-capital ratios, which Deng et al. already show predict large-cap crashes? This is the most serious threat. A firm with heavy short-term debt and little working capital is plausibly already a high-SFLI firm. SFLI differs in measuring a timing gap: long-term spending covered by short-term money. A firm can have high leverage that is all long-term, and so low SFLI. The ablation settles it directly. If the ratios absorb the effect, we will report that SFLI is redundant and that the right conclusion is "leverage structure conditions the size premium," a smaller but still new finding.

Why not just add a trading-liquidity factor? E2EAI finds trading liquidity negligible for large caps, which makes it a poor candidate for explaining large-cap stress divergence. Turiel and Aste, in Heterogeneous Criticality in High Frequency Finance, nonetheless show that the most liquid large caps dominate systemic intraday co-crashes. Trading liquidity could therefore matter in stress in a way the factor models miss. We include turnover-based illiquidity as a control in every cell. The shock-channel placebo further separates the two: trading-liquidity stories predict effects on both shock types, while the funding story predicts effects on funding shocks only.

Why not just feed SFLI into a deep factor model and read the attention weights? That would improve prediction without answering the economic question. Attention weights in DMFM and E2EAI are known to be hard to interpret, a problem noted alongside Deep Learning in Asset Pricing. A transparent double sort with an explicit placebo gives a falsifiable result. We still run the deep model as a secondary check, not as the main evidence.

Why not just use the five-factor investment factor? Gu, in Research on the Influence of Covid-19 on U.S. Oil Industry Based on Fama-French 5-Factor Model, found that size distinctions fade into profitability and investment preferences during recovery. SFLI could simply be aggressive investment under another name. The alphas in every cell are measured after controlling for the five factors, including investment. Investment growth is also in the ablation set.

Proposed Experiments

The first decisive check is the shock-channel contrast. The hypothesis is confirmed only if high-SFLI large caps show significantly negative five-factor abnormal returns relative to low-SFLI large caps in funding-stress months, with a near-zero difference on trade-news days. A similar penalty on both shock types would point to generic risk, not refinancing fragility, and we would reject the mismatch interpretation.

The second is the crisis decomposition. We measure how much of the pandemic-era shift toward large caps, the pattern documented by Zhu over 2017–2022, comes from the low-SFLI large-cap cell. The prediction is that most of it does.

The third is the crash-model comparison. We rerun the replicated Deng et al. classifier with SFLI added, then with SFLI replaced by the ablation controls. We report large- and small-cap accuracy, true-positive rate and false-positive rate across five seeds, alongside SHAP rankings. A large-cap gain without a small-cap gain would support the claim that mismatch matters where size is supposed to protect.

As a robustness check, we repeat the portfolio sorts using two alternative mismatch measures that need no cash-flow decomposition:

  • short-term debt share minus the fixed-asset share of total assets;
  • growth in receivables relative to short-term borrowing, the Sinopharm pattern.

This guards against results that hinge on one accounting formula.