To address this, three FTS variants - Chen’s original model, a heuristic high-order extension, and an FTS-MLP hybrid - are benchmarked alongside a recurrent neural network (RNN), the transformer-based PatchTST, and the channel-independent linear model NLinear. A total of 95 series (70 univariate, 25 multivariate) from the Towards Fair Benchmarking (TFB) suite, spanning ten domains and annotated for trend, seasonality, and stationarity, serve as the testbed.

Results indicate that the Hybrid FTS-MLP achieves the most balanced accuracy among fuzzy methods - particularly in volatile and multivariate contexts - while Chen’s model excels on strongly seasonal univariate series. Although PatchTST and NLinear maintain overall superiority, FTS approaches demonstrate competitiveness in niche scenarios. The study concludes with recommendations for automated hyper-parameter tuning, rule-base compression, and patch-style context windows to enhance FTS adaptability and narrow the remaining performance gap.