Time series data is hard, especially without lots of training examples. Consider looking into traditional forecasting methods as a baseline. AR(I)MA based appr

Time series data is hard, especially without lots of training examples. Consider looking into traditional forecasting methods as a baseline. AR(I)MA based approaches with seasonality might actually be fine for you. However, it all changes if there are structural breaks in your dataset. Look into Rob Hyndman's free online book (https://otexts.org/fpp2/ https://otexts.org/fpp2/) for an excellent introduction into timeseries forecasting.