Borrowed buying helped predict next-month returns in China from 2024 onward, leaving a short record for a trader to trust. Earlier signals said little; in 2023, they may have pointed the wrong way. Cui, Xiong and Wu pool those regimes in their headline estimates. The result offers a promising input for stock selection, backed by about two and a half years of evidence. It remains well short of a trading alpha.
Measuring borrowed buying
The authors divide each stock's monthly margin purchases by its monthly trading value. They call the ratio margin buying intensity. Its sample mean is 8.93%, with a median of 8.86%. They winsorize it at 1/99 each month, then standardize it across stocks. Every coefficient reported below therefore measures return per one cross-sectional standard deviation.
Financing costs supply the economic argument. A buyer using margin pays interest and bears margin-call risk, so the purchase may convey more conviction than a cash trade. Retail trend-chasing magnified by credit offers a competing explanation. The pilot-era studies cited in the paper disagree: Chang, Luo and Ren (2014) and Lv and Wu (2019) found no positive return prediction from margin trading. Yu et al. (2018) found a positive relation.
The sample covers Shanghai and Shenzhen A-shares and excludes the Beijing exchange. The raw data span January 2022 to June 2026. A year of return history is needed for the 12-2 momentum control, pushing the fully controlled sample's start to January 2023. For the one-month horizon, the authors have 131,462 stock-months across 3,574 firms and 41 months. They control for log size, log price, current-month return, 12-2 momentum and log turnover. They estimate firm and month fixed effects, run Fama-MacBeth regressions with Newey-West lags of 0, 2 and 5, and sort stocks into quintiles.
We could not test the signal ourselves. We lack both stock-level margin purchase amounts and the history of A-share eligibility. Price and volume cannot reconstruct a financing flow.
What do the high-minus-low returns show?
At one, three and six months, the panel coefficients are 0.0049 (t=8.71), 0.0083 (t=5.66) and 0.0096 (t=3.67). A one-SD increase predicts an additional 0.49, 0.83 and 0.96 points of cumulative return. The t-stats use firm-clustered errors across many stock-months. The authors warn against treating them as proof of a cross-sectional anomaly.
Fama-MacBeth provides the better gauge. Its margin buying intensity estimates are 0.0031 (t=2.58), 0.0069 (t=2.70) and 0.0088 (t=2.94). Each relies on just 36 to 41 monthly slopes.
The portfolio sorts buy the highest margin-buying-intensity quintile and short the lowest. The spreads look attractive: equal-weighted Q5-Q1 returns are 0.97% (t=2.52), 1.88% (t=2.24) and 3.66% (t=3.34). Value weighting raises them to 1.38%, 2.82% and 6.13%. Flow signals usually concentrate in small caps; these results run the other way.
Most of the one-month spread comes from the bottom quintile. Q1 returns 0.207%, whereas Q2 through Q5 return 0.911%, 1.063%, 1.154% and 1.176%. The jump from Q1 to Q2 accounts for roughly 0.70 of the 0.97-point spread. Q1 is also the portfolio a trader would need to short. The authors acknowledge that these are raw cumulative returns. They caution against reading the evidence as an immediately implementable strategy without accounting for transaction costs, constraints on shorting the low-intensity portfolio, and risk adjustment.
The break between 2023 and 2024
The abstract flags the change in sign. It deserves to lead the paper. In a regression confined to 2023, the one-month coefficient is -0.00347 (t=-3.21).
A fully interacted model allows every slope to change from 2024. For 2023, it estimates margin slopes of -0.0010 (t=-1.31), -0.0019 (t=-1.00) and -0.0064 (t=-1.94). From 2024 onward, the slopes are 0.0062, 0.0104 and 0.0143. The changes have t-stats of 9.27, 6.19 and 6.02. Positive evidence thus comes from signals formed between January 2024 and the last signal date for each horizon. For the one-month horizon, that date is May 2026.
The authors argue that time variation reconciles the conflicting pilot-era findings. Their conclusion describes the information content as state dependent and time varying, a reading consistent with the data. Yet the abstract and conclusion still give prominence to the pooled 0.49, 0.83 and 0.96 figures. Those figures average a negative-or-flat 2023 together with a positive period from 2024 onward.
The late-September 2024 rally is the obvious suspect. Excluding every signal with a return window touching September or October 2024 leaves coefficients of 0.0049 (t=8.52), 0.0090 and 0.0129 (t=4.46). The six-month estimate increases from the baseline 0.0096 to 0.0129. Within the period from 2024 onward, the three-month slope is also higher in non-up months (0.0184) than in up months (0.0069). The authors conclude that broad market direction alone cannot explain the shift.
They decline to identify a cause. That restraint is right. Possible explanations include the October 2022 expansion of eligibility and tighter securities lending during 2023 to 2024, culminating in the suspension of securities refinancing from July 2024. For a trader, the suspension has two consequences. It occurs within the period generating the effect, and it makes the Q1 short harder to establish.
Industry controls weaken the longer horizon
The abstract acknowledges greater sensitivity in the six-month estimate while saying that the short and medium horizons survive. They mostly do.
With industry-by-month fixed effects, the one-month coefficient is 0.0032 (t=5.80), and the three-month coefficient is 0.0041 (t=2.91). The six-month estimate drops to 0.0037 (t=1.48). Dividing purchases by market cap instead yields 0.0059 (t=4.89), 0.0078 (t=3.00) and 0.0031 (t=0.76). Industry-adjusted returns give a six-month estimate of -0.0024 (t=-0.80). The authors run that test only on a subset with incomplete classification coverage. The three-month result weakens there too, to t=1.20.
Only the one-month horizon survives industry-by-month fixed effects, industry-adjusted returns and market-cap scaling.
For the alternative financing measures, financing-balance growth misses at one month (t=0.87). The authors regard it as a slow-moving stock variable. The six-month spreads of 3.66% and 6.13% are the hardest headline figures to defend against the six-month checks of t=1.48, t=0.76 and t=-0.80 above.
The short-side signal is stronger than the long-side signal. Short-sale value divided by trading value produces coefficients of -0.0061 (t=-13.91), -0.0157 (t=-13.33) and -0.0270 (t=-12.68). Lendable supply and hedging demand influence that series, the authors caution. They also say they do not test its time variation. The evidence of instability therefore applies only to margin buying.
The authors do not estimate a multifactor alpha incorporating profitability, investment, quality, idiosyncratic volatility or lottery factors. We did not find returns after costs either. Exchange eligibility adds another dependency. Shenzhen added 400 names in October 2022; more than 70% had circulating market cap below RMB 10 billion. Only stocks on the exchanges' eligible lists can have this signal.
If post-2026 observations keep the one-month slope for the period from 2024 onward near 0.006, the signal earns a place in an A-share model. A return toward the 2023 estimates would show that the pooled average combined two regimes.