IPM Prediction result for Inconsistent historical data. - Accuracy is not good.

Summary:

We are trying to implement IPM insights and prediction. But we have many SKU-level series have long runs of zero periods followed by large spikes. For example, a SKU may show near-zero units for several months

Inconsistent series: Some intersections have very short or broken history because SKUs are introduced and retired frequently.

We have actuals for 36 months , but the prediction is not good. Getting around 45% only

Questions:

Which algorithm does Oracle recommend for intermittent and lumpy demand: AutoMLx, LightGBM, XGBoost, Prophet or SARIMAX? Are intermittent-demand methods such as Croston, SBA or TSB available, or on the roadmap?

Does Oracle recommend forecasting at an aggregated level (for example, product family) and then allocating to SKU, rather than predicting at the SKU level? If so what would be recommended approach.