A new research paper explores how to forecast a changing consumer preference while retaining an explicit representation of its vagueness. It combines a Mamba-type selective state-space recurrence with triangular fuzzy numbers, which describe a central value and a spread.
A constrained recurrent state
The model computes recurrence parameters from ordinary numerical features, while keeping the fuzzy part linear with non-negative coefficients. Under the stated construction, the authors show that the state remains a proper triangular number and a convex combination of past observations.
This design is intended to prevent fuzziness from accumulating indefinitely as the model updates. The paper also connects the algebraic operations to Zadeh's extension principle. These mathematical properties concern the defined recurrence, rather than proving that the model's forecasts will be reliable for every customer dataset.
What the experiment measures
The reported example is synthetic and includes competitor promotions and customer switching. A 107-parameter model achieves a forecast-center mean absolute error of 0.154, compared with 0.163 for a constant-inertia evolution operator and 0.179 for a naive forecast.
The model matches a crisp version of the same architecture while adding a linguistic interpretation of forecast vagueness. The authors also illustrate using that readout to identify a permanent switch to a competitor. These results should be read as a controlled demonstration; the paper does not establish real-world retail deployment performance from this experiment alone.
How to investigate the approach
A useful next evaluation would preserve time order on actual purchase data and compare the same baselines across different customer segments. Check whether the fuzzy readout helps a decision-maker understand an uncertain forecast, rather than assuming it is a calibrated probability interval. The contribution is a small, interpretable state representation with defined mathematical behavior, whose practical value still depends on the data and the decision it supports.
Comments