Indonesian bank study pairs sentiment t 42.9 with p 0.0916

The reported 98.42% fit for Indonesian bank returns needs an audit before anyone trades on it. Gombo, Wardhiah, Husaini and Zulfan attribute that share of return variation to news-based political sentiment and three national macro series.

What would the trade be?

The sample covers 40 banks on the Indonesia Stock Exchange (Bursa Efek Indonesia, IDX), selected from a population of 47 over 2020 to 2024. Returns are measured as plain capital gains, (Pt - Pt-1)/Pt-1. The regression has four inputs. Political sentiment comes from Kompas.com, Detik.com and CNN Indonesia; inflation measures the change in CPI. The policy rate is Bank Indonesia's (the central bank's) 7-Day Reverse Repo Rate. The exchange rate is BI's IDR/USD middle rate. The authors use a pooled panel regression, called the Common Effect Model, which gives every bank the same intercept and slope for each variable.

The economic argument is familiar. Stable politics and pro-market policy encourage demand for bank shares. Higher rates increase banks' funding costs and draw investors toward deposits and bonds. Inflation reduces purchasing power and compresses net interest margin. The authors expect a smaller role for the rupiah because Bank Indonesia's net open position rules limit banks' FX exposure.

The estimates have the expected signs: +1.535446 for sentiment, -0.036908 for inflation, -0.147994 for the policy rate and -0.000267 for the exchange rate. The paper calls all four significant and reports an F-statistic of 3,106.350.

A trader reading these as signals would buy Indonesian banks when sentiment is positive and policy rates are falling. Yet the equation places returns and all four regressors in the same period t. It therefore explains returns after the fact. We did not find a lagged specification or an out-of-sample test. Its forecasting value remains untested.

t-statistics of 43 beside p-values of 0.09: which column is right?

At the paper's 5% threshold, most of its reported p-values fail. Sentiment has 0.0916, inflation 0.0659 and the policy rate 0.0600. The exchange rate alone passes, at 0.0270. Both the abstract and conclusion nevertheless describe all four as significant.

The t-statistics tell a different story. The table gives sentiment a t of 42.9471, inflation -32.6078 and the rate -43.7668. With roughly 195 residual degrees of freedom, a t of 43 implies a p-value indistinguishable from zero.

Both columns cannot be right.

For three of the four slopes, dividing the coefficient by its standard error reproduces the t-statistic. Sentiment's 1.535446 divided by 0.035752 gives 42.95. The exchange-rate row breaks that relationship: -0.000267 divided by its standard error of 1.478039 gives about -0.0002. Its reported t is -18.1081. The only regressor that passes the p-value threshold thus has a t-statistic we cannot reconstruct.

The model-selection test suffers from the same contradiction. The Chow test determines whether to pool the banks or use bank fixed effects. Its F of 8.8767 on (39, 156) degrees of freedom exceeds any conventional critical value, yet the probability is reported as 0.6207. The chi-square version reports 10,290.80 on 39 degrees of freedom, with a probability of exactly 0.0500. Taken literally, the F statistic rejects pooling and with it the paper's chosen model. The authors instead use the 0.6207 to accept H0. Their methods also name a Hausman test and a random-effects model as candidates. We did not find either the Hausman result or a random-effects estimate.

The authors acknowledge a separate diagnostic failure. The Jarque-Bera probability is below 0.05, indicating residuals that are not strictly normal. They defend the result by pointing to residuals clustered near zero and skewness close to zero. That defence is fair within its limits. For a table with 200 observations, residual non-normality is the least of the inference problems.

Five annual periods, four regressors, one intercept

The sample's 200 observations come from 40 banks times 5 years. Each bank shares the same inflation, policy-rate and exchange-rate value within a period. On the annual frequency implied by the paper's 40 × 5 count, the macro slopes therefore have only five distinct time points behind them. We found no mention of standard errors clustered by period. The table appears to treat all 200 rows as independent.

Whether sentiment also has one value per year is left unclear. If it does, the model is saturated in time: an intercept and four slopes would fit five year-means exactly. Every bank would receive its year's average return as its fitted value. The coefficients would amount to an arbitrary relabelling of five numbers. An R-squared of 0.9845 would then tell us only that variation lies mostly between years. This interpretation depends on a data construction we could not inspect.

The reported magnitudes fit that possibility. Returns have a standard deviation of 0.1141 and range from -0.14 to 0.20. Multiplying sentiment's standard deviation, 0.2923, by its coefficient, 1.535446, gives about 0.45 of return. For the exchange rate, 841.08 times 0.000267 gives about 0.22. Each term moves more than the return itself. The regression therefore relies on those terms cancelling one another, just as near-collinear series with five data points would.

The sentiment measure is unresolved

The methods call sentiment a dummy, set to 1 when a political event or sentiment is present and 0 otherwise. The introduction instead describes an index of political news coverage. The descriptive statistics support the index interpretation: mean 0.034, median 0.14, minimum -0.42 and maximum 0.36. Negative values rule out a 0/1 variable, so the measure used was signed and fractional. We could not find the coding rules or establish its timing relative to returns.

The frequency is unclear too. Prices are said to come from idx.co.id and Yahoo Finance at monthly frequency. The observation count, however, is 40 firms × 5 years.

Three screens reduce the sample from 47 banks to 40. One requires complete data over 2020 to 2024, excluding any bank that listed or delisted during the period. That creates survivorship bias.

Before anyone sizes a position

We did not test the method because we could not. Our data contains no Indonesian equity prices, no Indonesian CPI, policy-rate or USD/IDR series, and no established historical coverage of the three news portals. Even if those inputs were available, the competing sentiment definitions and the monthly-versus-annual ambiguity leave no single procedure to replicate.

Four things would change our view: the sentiment coding rules, monthly returns with the signal lagged one period, standard errors clustered by time, and a table in which t, standard error and p agree. Until then, 98.42% remains a number to audit.