A trader cannot use the day's FPI total to get ahead of that day's NIFTY Bank move. The flow total arrives after the move, and its link to volatility fades under controls. In the paper's regression, daily net foreign flow has p =.163 after DII flow and lagged volatility enter. The authors acknowledge the weak result in their abstract, although they still describe daily FPI activity as one market signal among several. Selling gives even less reason for conviction: its unadjusted link to larger moves (r = -.172, p =.046) also loses significance under those controls (p =.163). Even if the link had survived, it would describe a same-day association with nothing to act on at the close.
What the study measures
The premise is familiar to anyone watching Indian flows. Foreign portfolio investors (FPI, the current SEBI category for what older work calls FIIs) supply liquidity, and concentrated withdrawals might move prices. The authors choose NIFTY Bank for the connection banks provide between credit, payments, savings and monetary transmission and the wider economy. They pair official NIFTY Bank closes with daily cash-market net flows for FPIs and domestic institutional investors (DIIs), sourced from the National Stock Exchange (NSE). An exact-date merge yields 135 trading days, from 14 January to 24 August 2026.
The volatility measure is absolute daily log return; squared return provides a check. The authors correlate FPI net flow with returns, then run an OLS of absolute return on FPI net, DII net and the previous day's absolute return. They also compare net-buying and net-selling days. Finally, they put buying and selling magnitudes in separate terms and test whether selling has the larger coefficient.
None of the four tests makes the case for a stronger selling effect. FPI net flow against return yields r =.114 (p =.189). In the volatility regression, the FPI coefficient is -.028 per ₹1,000 crore (p =.163); lagged absolute return is.179 (p =.040). The model explains 6.7% of the variance. The abstract gives recent volatility persistence more weight than net foreign flow in this short sample. I agree. The direct difference between the selling and buying coefficients is -.051.
The selling-day comparison
At first glance, selling days look choppier. Absolute returns average.904% across the 100 net-selling days, versus.764% across the 35 net-buying days. FPI net flow and absolute return have an unadjusted correlation of -.172 (p =.046): heavier selling accompanies bigger moves.
The group comparison is weaker. Its.140-point difference produces t(133) = -.870, p =.386, with Cohen's d at -.171. That uncontrolled comparison was never significant. The correlation is the result that loses significance once DII flow and yesterday's volatility enter: the FPI coefficient then has p =.163. Using squared returns takes it to p =.311.
FPIs sold net on 100 of 135 days. Across all 135, their average daily net flow was -₹2,377.05 crore, while DIIs averaged +₹2,872.66 crore. The series correlate at -.354. The authors see some domestic counter-positioning there, without claiming DIIs caused stability. Over the seven-month window, mean daily return was -0.072% and daily SD was 1.194.
Why does selling get the smaller p-value?
The asymmetric regression contains the paper's most revealing comparison. Selling magnitude has a coefficient of.047 (p =.044); buying magnitude has a coefficient of.098 (p =.231). Buying's point estimate is roughly double selling's. Selling gets the smaller p-value through a smaller standard error,.023 versus.082 for buying. The sample helps explain the gap: 100 selling days, including a net-flow low of -₹21,105.86 crore, against 35 buying days with a maximum of ₹4,864 crore. The greater number of selling days and wider spread of values shrink its standard error. That is all the p =.044 reflects.
The authors make the direct test, to their credit. Sell minus buy is -.051, with SE.078 and t(129) = -.658. The one-tailed p for stronger selling is.744. As their abstract says, the contrast did not show selling to be stronger than buying.
Their discussion calls the individually positive selling coefficient "broadly compatible" with crash-risk concerns, then says the failed contrast "prevents a strong asymmetry claim." The authors themselves stop short of that inference. My objection concerns a narrower line: they suggest regulators may continue monitoring concentrated sales, liquidity and investor-category composition while avoiding conclusions from a single-day flow number. Sales lead that list, although the contrast (one-tailed p =.744) gives no reason to put them ahead of purchases. The point estimates, if anything, face the other way.
The limits of this null
Flows may still matter. There are only 35 buying days, and the t-test treats days as independent despite the paper's finding of volatility persistence (p =.040). We did not find HAC or heteroskedasticity-consistent (White) standard errors anywhere.
The primary absolute-return model rejects residual normality (W =.890), and nine cases exceed the Cook's 4/N screen. Its Breusch-Pagan p is.118 and its Durbin-Watson is 2.102. HAC errors would not address non-normal residuals. The argument for them comes instead from the squared-return model's Breusch-Pagan p of.053 and the t-test's independence assumption.
The authors say a failure to reject "does not mean that the null statements were proven true." Agreed. Non-rejection gives no evidence of a zero effect; the significant selling coefficient gives no evidence of asymmetry. The FPI estimate (β = -.129) is about 70% as large as the lagged-volatility term (β =.180), but remains imprecise (SE.020, p =.163). I would change my view if the flow coefficient held up in a two-year sample using next-day timing and Newey-West errors.
Timing is the trading problem
Every flow coefficient pairs day-t net flow with day-t absolute return. The authors acknowledge "possible simultaneity": flows might respond to the move as readily as cause it. By the time the day's flow total is known, the day's move has occurred. Yesterday's absolute return is the sole regressor known in advance, and its.179 coefficient describes volatility clustering.
The flow data cover the cash market only, as the authors specify. Foreign positions in NIFTY Bank futures and options sit outside the study.
We could not test a next-day version ourselves. Our data lacks NIFTY Bank price history and daily FPI or DII cash-market flows. A US bank ETF would lose the foreign-versus-domestic institutional split at the heart of this question, so substituting one would discard the mechanism being tested.
The authors advise investors to use foreign-flow data as context rather than a stand-alone trading rule. These 135 days support that advice. And the -.051 contrast gives regulators no basis here to rank concentrated sales above purchases.