The four significant bank ratios in Arlan and Leon's study do not yet give investors a stock-selection signal. They track Indonesian bank stock moves over 2020 to 2024. Even the return series needs resolving: its maximum, 15.68632, cannot come from the formula the authors print.

What earned the investment advice?

The paper takes its starting point from Sar and Panigrahi's study of Indian banks. CAMELS is a supervisory scorecard for bank health. Because its ratios are public, signalling theory suggests investors should read them and reflect them in prices. The authors select six:

Leverage, defined as total debt over total assets, is the addition the authors present as novel. They control for size using total assets. Annual share price growth, labelled SPG, is the outcome.

The starting universe contains 47 IDX-listed banks. The exclusions are three with incomplete reports, four sharia banks, three regional banks and five with incomplete stock data. The remaining 32 banks supply 160 bank-years. The authors estimate a random-effects panel in EViews and assess it with t-tests at 5%.

Positive coefficients emerge for CAR (1.135214, p = 0.0073), CASA (0.343116, p = 0.0465) and leverage (11.93657, p = 0.0016). NIM has a negative coefficient of -10.49197 (p = 0.0134). NPA, ROCE, LATA and size are insignificant, with respective p-values of 0.6770, 0.7818, 0.4749 and 0.3858. The abstract translates those results into advice: investors "are also urged to take these signs into account when choosing which investments to make." This review disputes that recommendation.

The formula cannot reach 15.69

SPG is defined as (Pt − Pt−1)/Pt. Rewritten, that is 1 minus Pt−1/Pt, which stays below 1 whenever prices are positive.

Yet the maximum reported is 15.68632, recorded by Allo Bank (BBHI) in 2021.

Using the prior price as the denominator would make 15.69 a gain in which the stock multiplied about 16.7 times over a year. A 16.7-fold year is arithmetically possible; the printed definition cannot produce it. Bank Mayapada's minimum of -0.91373 in 2021 is compatible with either denominator and therefore cannot settle the question. The answer determines the scale of the coefficients.

The variable definitions raise other reconstruction problems. LATA is described as current assets over total assets, but reaches 1.066430. ASET carries the label total assets, although its mean, 31.81672, looks like a log. The fitted equation is attributed to Sar and Panigrahi's model even though it contains this paper's own coefficients. Each issue is minor on its own. Taken together, they prevent a reader from rebuilding the dataset from the variable table.

Four significant ratios, with no earlier date

Banks release their annual reports after year-end. The ratios come from those reports, yet the variable definitions specify no lag. We found no statement establishing that the ratios were measured before the return window began. Regressing a year's price change on that year's ratios captures a same-year association. You could not have traded on it. We raised this issue before with same-year ROA among LQ45 stocks.

The effects vary sharply when each coefficient is scaled by a standard deviation of its regressor. Leverage shifts SPG by about 1.09 (11.93657 times 0.091567), or roughly 109 points of price growth. For NIM, the shift is about -0.24 (-10.49197 times 0.023157). CAR produces about 0.19 (1.135214 times 0.164576). CASA contributes about 0.07 (0.343116 times 0.191344), with a p-value of 0.0465.

The positive signs on leverage and CAR are also uncomfortable companions. Debt over assets moves roughly against the equity share; CAR measures capital over risk-weighted assets. The model thus rewards both more debt per asset and more capital per risk-weighted asset. That combination calls for a collinearity check. We found none. The authors interpret leverage as investors seeing "more debt as a way to boost the bank's ability to distribute credit". A reported variance inflation factor would make that explanation easier to accept.

The authors say the estimation addresses heteroscedasticity, autocorrelation and cross-sectional dependence. Readers get coefficients and p-values in the results table. We found no R², standard errors or Hausman statistic.

BBHI supplies five of nine extremes

At -0.057225, median SPG describes a bank-year that lost ground. Mean SPG is a gain of 0.262819, with a standard deviation of 1.659117. BBHI's 2021 observation contributes 15.68632/160, about 0.098, to the mean by itself. That single observation accounts for more than a third of the 0.262819 mean.

BBHI also supplies the minimum for four regressors: CASA (0.079030), NPA (0.000125), size (28.58139) and leverage (0.420260). The estimation fits nine parameters to 160 rows. A bank occupying these extremes in both the outcome and the regressors warrants a run excluding BBHI and another with winsorized SPG. We found neither. Nor did we find year effects, despite a window beginning in the 2020 pandemic year and containing BBHI's 15.69 jump in 2021.

Survivorship enters through the first selection rule: banks must remain continuously listed on IDX through 2020 to 2024. Another five banks disappear because their stock data are incomplete. The surviving sample of 32 banks therefore consists of those that lasted throughout the period.

Keep CAMELS as a hypothesis. Keep the capital back.

The conclusion proposes adding macroeconomic indicators, corporate governance and banking digitization, as well as studying longer periods. Those extensions are reasonable. The immediate work is to recompute SPG with the denominator actually used and lag the ratios to their publication dates. A test on years outside 2020 to 2024 becomes worth running only after those steps.

We could not run these checks ourselves. Our data covers US equities and contains neither IDX prices nor Indonesian bank filings. US bank fundamentals cannot substitute for Indonesian capital measured against risk-weighted assets or CASA based on Indonesian deposit composition. Using generic balance-sheet ratios would test a different hypothesis.

A result that would change my view is straightforward. Lag the ratios to publication, remove BBHI, and show leverage retaining a positive coefficient at anything like 11.94 across the 31 remaining banks. Then the paper has something. Until that run exists, its four coefficients describe what happened to these banks over 2020 to 2024. We found no lagged or out-of-sample test in the paper showing that they predict returns.