Dhaka's data cannot settle whether August 2024 changed efficiency

Daily returns on Dhaka's largest stocks show serial dependence, with almost none of it looking like profit a trader could collect.

Sarwar, Sultana, Peeya and Rishad ask whether the DSE30 Index and its 30 constituents follow a random walk, and whether the government's fall on 5 August 2024 changed that. Their daily closing prices run from 4 January 2016 to 14 August 2025, giving roughly 2,165 to 2,275 observations per series. They apply five standard tests. Ljung-Box at 5, 10 and 20 lags checks dependence, alongside a runs test on return signs. Predictability is tested with the Lo-MacKinlay variance ratio at q = 2, 5 and 10 and an AR(1) with Newey-West errors. ADF (augmented Dickey-Fuller unit-root tests) is applied to prices and returns alike.

They repeat everything for a post-transition window beginning 11 August 2024. It covers about 240 trading days, excluding the 6 to 8 August sessions in which the DSEX gained 13.3%. A Chow test checks for a break in the AR(1) parameters at the transition.

Full-sample rejections are widespread. Across the three lag lengths, Ljung-Box rejects for 26, 30 and 28 of 31 series; the runs test rejects for 29 of 31. Significant AR(1) slopes appear in 13 of 31, including the index at p < 0.001. VR(2) rejects for the same 13. Rejections fall sharply after the transition: Ljung-Box drops to 9.7% to 29.0%, runs to 2 of 31 and AR(1) to 3 of 31. The authors resist interpreting this as improved efficiency. They point instead to a rise in the median AR(1) coefficient from 0.055 to 0.080, larger post-break coefficients in 18 of 31 series, and joint Chow rejections in only 4 of 31. Their conclusion is that "a change in government alone has not improved price efficiency."

We could not test any of this ourselves. We have neither Bangladesh-listed prices nor the DSE30 index history, and we lack a constituent roster for the period. Another market would tell us nothing about floor prices or a political shock in Dhaka, so we made no substitution.

Can a trader collect the serial dependence?

Accept the statistics, then examine their source. Every stock has a zero median daily return except Heidelberg Cement, LafargeHolcim and Olympic Industries. That means many days with no return. Floor prices constrained most listed stocks from March 2020 and again from July 2022. The authors themselves say these floors "mechanically" produce serial dependence and non-random runs. We did not find an account of how the runs test treats zero returns. With zero medians in 27 of 30 stocks, its 93.5% rejection rate could largely be counting stale prints.

The index raises a similar problem. The DSE30 has a VR(10) of 1.46. Citing Lo and MacKinlay (1990), the authors attribute index-level dependence to constituents trading at different times. If that explains the result, the autocorrelation belongs to a price at which nobody could trade.

The coefficients are small, too. Use the pre-transition median of 0.055 with a daily standard deviation near 2%, typical of constituents in the descriptive table. That gives a next-day forecast of about 11 basis points, before spreads and the ±10% limit. The authors describe the coefficients as "economically small"; they test neither a strategy nor its costs. Their own sentence makes the point: "Statistical predictability is not synonymous with an exploitable profit opportunity."

The roster adds another caution. They take DSE30 membership as of 14 August 2025 and apply it back to January 2016. Selecting firms that ended the period large and liquid produces a panel different from the one a trader faced in 2017.

A shorter window can explain fewer rejections

The authors are right about statistical power. Under the null, the standard error of a lag-one autocorrelation is roughly one over the square root of sample size. Roughly 2,000 pre-transition observations put that error near 0.022. A coefficient of 0.055 is then around 2.5 standard errors from zero. At 240 observations, the error is near 0.065, leaving even 0.080 barely more than one standard error away. The paper's Newey-West errors are wider still: only 13 of 31 full-sample slopes clear 5%. An effect this small should lose rejections when the window shrinks.

That leaves the before and after coefficients carrying much of the conclusion. They are the first of its two supports; the Chow test is the second. Yet the paper says the coefficients are "not tabulated." The median's move from 0.055 to 0.080 is smaller than one post-transition standard error. It can no more establish stable coefficients than the rejection counts can establish falling ones.

Only 4 of 31 Chow tests detect an August 2024 break

Thirty-one tests at 5% would be expected to yield about 1.5 false rejections. The four joint rejections (ACI Formulation, BATB, Beximco, National Bank) and three slope rejections are close to that. We found no multiple-testing adjustment. The index also fails to reject, with joint p = 0.713 and slope p = 0.412.

The authors interpret this as an absence of a systematic break. But the Chow test uses the same 240 post-break days whose low power they invoke to dismiss the decline in rejection rates. That lack of power cannot dismiss an efficiency gain and certify stability at the same time. On that reading, the Chow test could not detect a change of 0.025 in either direction.

Much of the paper's body acknowledges the problem. The comparison section says the post-transition evidence "does not yet establish an improvement in efficiency"; the conclusion says it "does not show" that the transition changed anything. The abstract goes further with "has not improved", and the conclusion's regulator paragraph repeats that stronger claim. Why does the abstract claim more than the body can support?

The four rejections offer little help. The authors note extreme-return days or floor prices in several rejecting series. Two of the four, BATB and Beximco, have the very extreme-return days or floor prices they flag. BATB records daily log returns of -111.13 and 109.86, and kurtosis of 465.14. The authors ascribe these to unadjusted corporate actions and identify them as a possible source of spurious negative autocorrelation. They retain them. Beximco remains at a floor throughout the subsample, leaving 165 observations. Meanwhile, the 27 non-rejections still depend on about 240 post-break days.

Raw prices, floors and the full-sample evidence

The authors deserve credit for listing raw prices, the DSE30-only universe and the short window among their limitations. Those same limitations make the full-sample result difficult to put to work on a desk. Contaminated prints swell the tail statistics: BATB's kurtosis reaches 465.14, compared with 16.49 for the index. Floors and the ±10% limit force continuation by rule, revealing nothing about the speed at which traders absorb news. The index also has only 2,165 observations, versus more than 2,200 for most constituents. The sole explanation is "data availability."

One index result survives the short window: the runs test still rejects after the transition at p = 0.007. The authors highlight it as well, calling for another look when a longer post-transition series is available.

The trader's question remains untested: does dependence survive on adjusted prices, with floor days removed and a realistic Dhaka spread deducted? The authors leave the adjusted-price rerun to future work. Until then, the paper documents frictions, consistent with the authors' own interpretation of their finding.

Its political question remains open.