A large fraction of computer and data science forecasting workloads involve discrete events and counts (clicks, orders, incidents), for which Gaussian emission models are inappropriate. We introduce Stanza-GLM, a nonlinear state-space framework that retains the interpretability and calibrated multi-horizon uncertainty of Stanza-like latent dynamics while generalizing the observation model to the exponential family (Poisson, Negative Binomial, Bernoulli). We develop efficient filtering and smoothing via iterated extended Kalman updates or Laplace moment matching, yielding reliable predictive intervals across horizons with optional conformal calibration. Across event- and count-heavy benchmarks, Stanza-GLM reduces predictive deviance and improves empirical coverage versus Gaussian emissions and deep sequence baselines while maintaining competitive runtime. Our ablations dissect the contribution of dynamic lag weights, dispersion, and inference choices, providing a practical recipe for production deployment in discrete-event time series.
A large fraction of computer and data science forecasting workloads involve discrete events and counts (clicks, orders, incidents), for which Gaussian emission models are inappropriate. We introduce Stanza-GLM, a nonlinear state-space framework that retains the interpretability and calibrated multi-horizon uncertainty of Stanza-like latent dynamics while generalizing the observation model to the exponential family (Poisson, Negative Binomial, Bernoulli). We develop efficient filtering and smoothing via iterated extended Kalman updates or Laplace moment matching, yielding reliable predictive intervals across horizons with optional conformal calibration. Across event- and count-heavy benchmarks, Stanza-GLM reduces predictive deviance and improves empirical coverage versus Gaussian emissions and deep sequence baselines while maintaining competitive runtime. Our ablations dissect the contribution of dynamic lag weights, dispersion, and inference choices, providing a practical recipe for production deployment in discrete-event time series.
