LoginSign Up
← Back to all sparks
E

epinowcast

ANALYTICS
Velocity0.0

A Bayesian Framework for Real-time Infectious Disease Surveillance

github.com/epinowcast/epinowcast ↗

epinowcast added Gaussian processes to its formula interface and made the sampler twice as fast

rbayesianstanepidemiologynowcastinggaussian process
◆Current state
epinowcast is a Bayesian nowcasting toolkit for right-truncated epidemiological count data, built on Stan with a brms-style formula interface. Over 2025-2026 it moved from experimental to stable, prepared for CRAN, and broadened past nowcasting proper — 0.6.0's max_delay = 1 support allows purely retrospective fitting of fully reported counts. 0.7.0 in July 2026 is the largest modelling release in the window.
◆Where it's heading
The package is converging on a general formula-driven latent process toolkit rather than a single nowcasting model. rw() and arima() were joined in 0.7.0 by gp(), a Hilbert-space reduced-rank Gaussian process placeable on any module's linear predictor with selectable kernels and an integration order matching arima()'s d. Alongside it, the fixed-effects design and integrated residuals are now centred against the module intercept, which the notes report roughly doubles sampling speed on a weekly random-walk growth model.
◆Prediction
With CRAN preparation done in 0.6.0 and the model surface substantially widened in 0.7.0, the next release is likely a CRAN submission plus consolidation of the gp() kernels. The release notes repeatedly benchmark against EpiNow2's behaviour, suggesting continued convergence between the two codebases.

◆Recent moves

  1. 1mo ago

    gp() brings Gaussian processes to every module, and centring doubles sampling speed

    ⚡ SPARK

    A gp() formula helper places an approximate Gaussian process on any module's linear predictor — growth rate, latent-to-observation proportion, both delay forms, report hazards and missing-reference proportion — using a Hilbert-space reduced-rank approximation with five selectable kernels. Separately, the fixed-effects design and integrated residuals are centred against the module intercept, decorrelating intercept from slopes and latent drift.

    View source ↗
  2. 3mo ago

    CRAN preparation, inspection methods, and retrospective-only fitting

    Moves cmdstanr and scoringutils to Suggests, adds print(), summary(), plot() and enw_get_data() accessors, and restructures the getting-started vignette around three competing models to present the package as a toolkit rather than one model. max_delay = 1 support means it can now fit fully reported counts with no nowcasting at all, widening the use case beyond right-truncation.

    View source ↗
  3. 7mo ago

    Structural reporting patterns for non-daily reporting cycles

    Adds enw_structural_reporting_metadata() and enw_dayofweek_structural_reporting() so weekly or other non-daily reporting cycles can be modelled over a daily generative process, with precomputed sparse index lookups for the aggregation. Closes a gap the features vignette had listed as in development.

    View source ↗
  4. 10mo ago

    Lifecycle moves to stable; pathfinder and negative binomial support

    The package declares interface stability, adds experimental pathfinder support for rapid prototyping and HMC initialisation, and introduces a negative binomial observation model with a linear mean-variance relationship. Sparse design matrices address memory pressure on larger models. The stability declaration is the notable part — it signals the formula interface is now something to build on.

    View source ↗
  5. 2y ago

    Non-parametric reference date models and a stricter max_delay default

    Adds support for non-parametric reference date models alongside the existing parametric form, and changes the max_delay default to the maximum observed delay with a warning, pushing users to set it from their data. Broad contributor list on this release relative to the ones around it.

    View source ↗
  6. 3y ago

    Fix initial conditions for cmdstan 2.32.1 compatibility

    Passes optional initial conditions as arrays to satisfy a cmdstan 2.32.1 change, plus input checking on max_delay. Required to use newer cmdstan, but no modelling change.

    View source ↗