frpp combines exponential kernels from emhawkes with a user-specified residual distribution. It provides model specification, simulation, likelihood evaluation and maximum likelihood estimation.
This is a development package. Its model object contains a kernel and a residual law; it does not inherit from the Hawkes specification.
Start here
- Get started: installation, a first model and simulation.
- Examples: one- and two-dimensional models migrated from emhawkes.
- Legacy emhawkes examples: the complete original FRPP examples and fitting code.
- Fitting and inference: optimizers, identification, diagnostics and restoration.
- 10,000-row fit validation: matched original and migrated estimates, curvature and elapsed times.
- Model and computation: residual refresh rules, kernel mapping and the emhawkes boundary.
- Function reference: arguments and return values.
Main functions
| Task | Function |
|---|---|
| Specify a residual law | residual_law() |
| Compose a model | fspec() |
| Simulate a path |
fsim() or stats::simulate()
|
| Evaluate the likelihood | logLik() |
| Fit parameters | ffit() |
| Inspect a fit |
summary(), coef(), vcov(), fdiagnostics()
|
| Restore a fitted model |
fitted_fspec() and finitial_state()
|
Simulation redraws every dimension’s candidate residual after every event in the combined process. This is a model rule, not an independent renewal clock for each dimension.
Current implementation
The default likelihood backend and simulator use public emhawkes interfaces. The experimental state backend assembles the FRPP likelihood from a shared kernel replay API available in GitHub development emhawkes 1.0.1.9000. Optional state simulation owns the FRPP event loop and uses event-wise kernel APIs from current GitHub development emhawkes. The bridge remains the default because the numeric reference loop is slower for large paths. Examples detect available APIs and label the simulation backend and comparisons.
Residual parameters remain fixed during fitting unless joint estimation is requested. A successful optimizer or a positive covariance matrix does not establish valid confidence-interval coverage. The guides explain these limits.