Accurate and timely forecasting of solar flares is crucial for mitigating their adverse effects. This paper demonstrates the significant predictive capability of low-dimensional microwave spectral representations from the RATAN-600 radio telescope for solar flare forecasting, offering a promising alternative or complementary data source. To rigorously evaluate this, we benchmarked radio-data-driven models against those using established photospheric magnetic field parameters (HMI SHARPs) within an adaptive machine learning framework. This framework utilized prequential (interleaved test-then-train) evaluation to simulate real-time forecasting and ensure temporally causal validation, which is often overlooked in conventional methods. Both data streams informed identical hierarchical multi-target classifiers predicting C, M, and X-class flare probabilities for 0–24h and 24–48h windows. Although SHARP-based models have traditionally been seen as strong performers, our experiments showed that models trained on RATAN-600 radio data not only matched, but in some cases slightly outperformed SHARP-based models across several evaluation metrics, including the True Skill Statistic (TSS), Probability of Detection (POD), False Alarm Rate (FAR), and Brier score. Both model types consistently outperformed randomized predictors by a wide margin in terms of prequential loss, although their performance exhibited notable variability aligned with the solar cycle. The prequential analysis underscored the sensitivity of both approaches to evolving solar conditions (assessed up to the rising phase of Solar Cycle 25 in 2025), emphasizing the importance of continuous, adaptive validation in developing robust operational solar flare forecasting systems. These results also suggest that integrating multi-modal data, including promising radio-derived features, could further enhance predictive capabilities.