This paper offers no tradable signal yet. Its positive ROA coefficient pairs annual returns with ROA from the same year, using 18 stocks selected for having survived. Whether investors knew that ROA before earning the return hinges on a price window the paper leaves unspecified.
Saputri and Septiani examine Indonesia's large-cap index in "The effect of market anomaly, financial performance and firm size on stock returns" (Pengaruh Anomali Pasar, Kinerja Keuangan dan Ukuran Perusahaan terhadap Return Saham). They ask whether profitability, firm size and a calendar pattern explain returns among LQ45 members. Their argument draws on two theories. Signalling theory says high return on assets shows investors that management uses assets well, drawing demand for the stock. The efficient market hypothesis says a recurring seasonal pattern should disappear. The authors predict positive effects from the anomaly (H2), ROA (H3) and size (H4), alongside a joint effect (H1).
Their annual panel covers 18 firms over five years, 2020-2024, yielding 90 firm-years. Returns measure capital gains alone, (Pt - Pt-1)/Pt-1, excluding dividends. ROA divides net income by total assets; size is ln(total assets). The authors label their "market anomaly" regressor a month-of-the-year effect. Chow and Hausman tests favour fixed effects, giving each firm a separate intercept (Hausman chi-square 20.014826, 3 d.f., p 0.0002).
The ROA coefficient is 3.129778 (t 4.62, p 0.0000). Size comes in at -0.355114 (t -2.05, p 0.0445), defeating H4 on sign. For the anomaly, the estimate is -0.017392 (p 0.6985). The joint F has p 0.002795; adjusted R² reaches 0.250067. The authors acknowledge three limitations: a short period, a sample restricted to LQ45, and 75% of variation left unexplained.
Survival determines the sample
The starting universe contains 68 firms with LQ45 membership at some point during 2020-2024. Removing 43 firms without continuous membership leaves 25. Excluding another 7 that report in currencies other than rupiah reduces the sample to 18. LQ45 reviews membership twice a year, considering market capitalisation, transaction value and financial health. Any stock that lost its place at a review during those five years is excluded. Selection thus depends on outcomes unavailable at the beginning of each return year.
For position sizing, the fixed effects deserve particular attention. Each of the 18 firms gets its own intercept. The slopes compare better and worse years within the same firm. They offer no guide to which firm to own.
Can this regressor test a calendar pattern?
The paper states its aim explicitly: "Pada penelitian ini anomali yang akan digunakan ialah Month of The Year Effect." Such a test compares average returns across calendar months, usually through month dummies and a joint test. Instead, the authors average each stock's monthly returns within a year: R̄ = (R1+...+Rn)/n. That produces a single observation per firm-year. Twelve months become one figure, calculated from the same prices used for the dependent variable. January and October are no longer distinguishable.
We could not reconcile the units either. The proxy's mean is 0.734759 and its maximum is 2.004167. Annual returns have a mean of -0.032651 and a maximum of 1.303571. Interpreting the proxy as decimals implies an average monthly gain of 73%.
The discussion nevertheless says a month of the year effect "terjadi di pasar modal Indonesia" (occurs in the Indonesian market). It describes significant differences in average monthly returns and attributes them to political, economic, cultural and holiday factors. Elsewhere in that same section, the authors treat the coefficient's p of 0.6985 as evidence for market efficiency. A regressor without calendar months supports neither interpretation.
Reading the ROA and asset-growth coefficients
The ROA estimate is large. Its pooled standard deviation is 0.078575; multiplying that by 3.129778 produces about 0.246, or roughly 25 points of annual return. The standard deviation of returns is 0.251643. But the pooled dispersion exceeds the within-firm changes used by fixed effects. Treat those 25 points as a ceiling.
Size takes more unpacking. Firm intercepts absorb differences in asset levels, leaving ln(total assets) to change within firms as assets grow. With a coefficient of -0.355114, asset growth of 10% (0.0953 in logs) accompanies about 3.4 points lower return in that year. This measures an association with asset growth in book terms. It says nothing about small versus large market capitalisation. Across the pooled sample, ROA and size have a correlation of -0.611222. Fixed effects strip out most of that cross-sectional variation.
Timing remains the central obstacle. Firms must have fiscal years ending 31 December to enter the sample. We did not find any statement specifying the opening and closing dates of the price windows. Under a calendar-year window, year-t ROA becomes public only after investors have earned the return paired with it.
What supports the reported fit?
The paper interprets its F as a joint test of the three regressors. It also tests the 17 firm intercepts. Including 17 firm dummies alongside 3 regressors gives the F 20 numerator restrictions and 69 residual degrees of freedom. The calculation (0.418591/20)/(0.581409/69) yields 2.48, matching the reported 2.483861. Adjusted R² follows the same arithmetic: 1 - 0.581409 × 89/69 = 0.250. The headline 25% therefore describes in-sample fit with a separate intercept for every stock. It has no out-of-sample test behind it.
The autocorrelation conclusion needs correction too. At 2.356473, the Durbin-Watson statistic lies between 4-dU (2.2736) and 4-dL (2.4111), the inconclusive zone. The paper declares the data free of autocorrelation after checking that DW lies between dU and 4-dL. That inequality holds. It uses the wrong band: the clean band is dU < DW < 4-dU. Durbin-Watson is a time-series test in any case, with only five years available for each firm here.
We did not find clustered or heteroskedasticity-consistent standard errors. ROA has a Glejser p of 0.0731, while return kurtosis is 11.11280. The effects tests address cross-sections alone. A common shock in 2020 or 2024 consequently remains in the residual.
The size finding hangs on p 0.0445.
We could not run this ourselves. We lack Indonesia Stock Exchange (IDX) prices and fundamentals, as well as historical LQ45 membership. Reconstructing the 68-to-18 filter requires that membership history. So does testing the 43 firms removed by the continuity requirement.
We would take seriously a cross-sectional sort on ROA lagged beyond the annual report date, covering all 68 ever-constituents and earning a positive spread after costs.