KARMA Explains Multivariate Time Series Forecasts

Many existing explainable AI (XAI) methods are not specifically designed for time series forecasting and often incorrectly assume independence between timestamp features. This oversight ignores the fundamental temporal dependencies in time series data, leading to explanations that can misrepresent the sequential and causal structure. Researchers have developed KARMA (K-Order Markov Approximations), a new XAI method tailored for time series predictors. KARMA addresses the temporal dependence issue by constructing a Markov surrogate model that accurately reflects the dependencies learned by the original predictor. The approach involves three key steps: first, identifying the optimal history length (K) that is sufficient for the model's predictions; second, estimating the best-fitting K-order Markov transition kernel from the discretized history space; and third, deriving a comprehensive five-level global explanation hierarchy from this kernel. Demonstrated on real-world weather data and complex synthetic data, KARMA effectively recovers causal structures and identifies temporal dependencies more accurately than traditional attribution methods like TimeSHAP.