Note(s): ***p < 0.01, **p < 0.05, *p < 0.10. Standard errors in parentheses. Column 1 uses the conventional rolling volatility SD(ROA_{t-2}, ROA_{t-1}, ROA_t), which shares the year t-1 ROA observation with the ESG regressor and is reported solely for comparability with prior literature. Columns 2 and 3 are the preferred specifications using clean backward-looking windows that exclude the current year, eliminating any mechanical overlap between the dependent variable and the ESG regressor. A diagnostic comparison confirms that the overlap in Column 1 attenuates rather than inflates the main coefficients: the S_dev bias is −5.8 per cent and the G_dev bias is 41.7 per cent, both in the direction of understatement. Column 4 tests whether current ESG signals predict future earnings instability