Whether different improvement mechanisms in grey forecasting models can provide stable out-of-sample forecasting gains under small-sample conditions requires evaluation under consistent settings. Using China’s national average overall digital finance index from 2011 to 2023, this study develops a unified evaluation framework to compare GM(1,1), the GM(1,1) power model (GPM), the conformable fractional GM(1,1) power model (CFGPM), and three CFGPM extensions based on initial-condition optimization, background-value optimization, and logarithmic transformation. The models are evaluated using a rolling-origin expanding-window approach, sensitivity analysis of the initial estimation window, and the exact Wilcoxon signed-rank test. The out-of-sample MAPEs of GM(1,1), GPM, and CFGPM are 12.08%, 3.49%, and 2.96%, respectively. The CFGPM also has a lower Median APE than the GPM, although their paired forecast-error difference is not statistically significant. In contrast, the difference between the CFGPM and GM(1,1) is significant at the 5% level. The three additional optimization strategies reduce errors at some forecast origins but provide neither a stable overall advantage over the CFGPM nor a statistically significant improvement. The sensitivity analysis gives the same main comparative results. Overall, the power-function structure and conformable fractional accumulation are associated with lower out-of-sample forecast errors, while further optimization does not necessarily produce stable forecasting gains. Re-estimating the CFGPM with all observations from 2011 to 2023 shows that China’s national average overall digital finance index is expected to continue increasing from 2024 to 2027, with gradually smaller annual increments.