Fit a Hawkes process by maximum likelihood (analytic gradient, fast routes)

Fits a self-exciting (Hawkes) point process to a series of event times, such as the times of shootings, break-ins or calls for service, and returns the estimated background rate, the branching ratio (the share of events triggered by earlier events) and the shape of the triggering kernel, with AIC, BIC and a goodness-of-fit test. Use it to ask whether events cluster in time beyond what a steady or seasonal rate explains.

Usage

chwkfit(
  times,
  hor,
  kernel = c("exponential", "weibull", "gamma", "lomax"),
  base = c("constant", "sinusoidal"),
  meth = "auto",
  eps = 1e-09,
  start = NULL
)

Arguments

times: Numeric vector of event times, sorted increasingly, all in [0, horizon], in any time unit (days are the natural unit for the sinusoidal baseline's yearly cycle). At least 50 events.
hor: End of the observation window, one positive number in the same unit as times; the window is [0, horizon]. With the sinusoidal baseline (an annual cycle, times in days) at most 4194304 days: that baseline is integrated a day at a time.
kernel: The triggering kernel: "exponential" (the default), "weibull", "gamma" or "lomax".
base: "constant" (the default) or "sinusoidal" (trend plus a 365.25-unit cycle).
meth: One of "auto" (the default), "exact", "soe", "truncate", "em", "inar". "auto" is "exact" for the exponential kernel, "truncate" for Weibull and "soe" for Lomax and gamma.
eps: Error level, one number in (0, 1), default 1e-9: the relative error per intensity of "soe" and the kernel tail mass dropped by "truncate". Ignored by the other routes.
start: Optional numeric vector of starting values in the order of theta: the baseline terms (a0, or a0..a3), then eta, then the kernel parameters. Values outside the parameter box are moved onto it. NULL (the default) uses a start from the data.

Value

A list:

theta: numeric, all estimates in the order baseline, eta, kernel.
basepar: numeric, a0 (constant) or a0..a3 (sinusoidal), on the log scale.
braratio: numeric, $\eta$, the expected number of events each event triggers.
kerpar: numeric, the kernel parameters (beta; alpha, lambda; alpha, beta; or alpha, c).
nll, aic, bic: numeric: the exact negative log-likelihood, $2k + 2\,nll$ and $k \log n + 2\,nll$.
n, horizon, kpar: the number of events (integer), the window end and the number of parameters (integer).
ksstat, kspva: numeric, the Kolmogorov-Smirnov test of the time-rescaling residuals against the uniform.
kernel, baseline, method: character, the choices used (method after "auto" is resolved).
eps: numeric, eps for "soe" and "truncate", NA otherwise.
converged: logical, the optimiser stopped at a maximum within the parameter box (before its iteration limit, at a finite likelihood). A maximum on the box counts: atbound names the parameters that lie on it.
atbound: character, the names of parameters that lie on the box (character(0) for an interior maximum).

A warning is given when $\eta$ reaches its upper bound 0.99 (a near-critical, probably non-stationary fit). An estimate on the box is not an interior maximum of the likelihood: day-dated (tied) times drive a shape or decay to its wall, and a Lomax fitted to exponential data runs to its exponential limit.

Details

The model is the univariate Hawkes process $$\lambda(t) = \nu(t) + \eta \sum_{t_j < t} g(t - t_j)$$ with a constant baseline ($\nu = e^{a_0}$) or a sinusoidal one ($\nu(t) = \exp(a_0 + a_1 t/T + a_2 \sin(2\pi t/365.25) + a_3 \cos(2\pi t/365.25))$, a trend plus a yearly cycle when times are in days, its integral by the trapezoid rule on max(64, floor(T) + 1) points), branching ratio $\eta \in [0.001, 0.99]$ and one of four normalised kernels $g$: exponential (decay beta), Weibull (shape alpha, scale lambda), gamma (shape alpha, rate beta) or Lomax (shape alpha, scale c). The log-likelihood is $\sum_i \log \lambda(t_i) - \int_0^T \lambda(t)\,dt$. Every route uses the analytic gradient of the shared C++ core (the code morie's Python arm calls), so one likelihood evaluation per point instead of one per parameter. The optimiser is a projected BFGS inside a fixed parameter box ($a_0 \in [-15, 15]$, the other baseline terms in $[-5, 5]$, kernel parameters bounded as listed in the source), at most 2000 iterations.

"exact": Ozaki's O(n) recursion for the exponential kernel; for Weibull and gamma the double sum stops where the kernel underflows to exactly 0 (the same value as the full sum); Lomax the full O(n^2) sum.
"soe": the completely monotone kernels (Lomax; gamma with shape < 1) as a sum of exponentials (Beylkin & Monzon 2010): relative error eps on every intensity, O(n K); a gamma kernel with shape >= 1 is truncated at eps.
"truncate": each event excites only lags with kernel tail mass above eps: an approximation for light-tailed kernels, O(n w).
"em": the EM algorithm (Veen & Schoenberg 2008), finished by the projected BFGS on the exact likelihood from EM's point (EM's steps shrink before it reaches the maximum). It is an R loop of up to 500 iterations and much slower than the other routes (tens of seconds for a few hundred events); because it starts from the same point but moves differently, it can end at a different local maximum.
"inar": Kirchner's (2017) INAR(p) least-squares estimator on binned counts (constant baseline only): a different, fast, approximate estimator.
"auto": "exact" for the exponential kernel, "truncate" for Weibull, "soe" for gamma and Lomax.

For the Weibull, gamma and Lomax kernels without start, a second deterministic start at the exponential fit (the shape-1 member) is also run and the better optimum kept, because the likelihood is multimodal. The exponential kernel without start is also run from a second start (branching ratio 0.1 and a decay a quarter of the reciprocal mean gap), and the better optimum kept: on data with little clustering its likelihood can have more than one maximum, one of them with the decay on its bound. A parameter on its bound (see atbound) is still a reason to refit from another start and compare nll.

The reported nll is the exact likelihood at the estimate whatever the route, so AIC is comparable across routes. An event exactly at hor is part of the record and counted, as in morie's Python fit. The Kolmogorov-Smirnov test compares the time-rescaling residuals of chwkres() with the uniform distribution (exact p-value up to 10,000 events without ties, asymptotic otherwise).

Events at the same instant do not excite each other: the intensity sums over strictly earlier events, as the definition of a Hawkes process (a simple point process, with no two events at one time) has it. Tied times are fitted that way, with a warning: data recorded to a resolution, such as daily dates, carry ties the process itself never makes, and treating them as exact biases the fit (Filimonov & Sornette 2015), so spread them uniformly across their interval first with chwkjit().

Errors: times with NA, unsorted, or outside [0, horizon]; hor not one positive finite number; fewer than 50 events; an event rate n / horizon outside $[e^{-15}, 100 e^{15}]$ per unit time (rescale the times); eps not in (0, 1); start of the wrong length or with a non-finite likelihood; method = "soe" with a kernel other than Lomax or gamma; method = "inar" with the sinusoidal baseline. Long fits show a progress spinner (EM a bar) on stderr unless options(morie.quiet = TRUE).

References

Ozaki T (1979). Maximum likelihood estimation of Hawkes' self-exciting point processes. Ann. Inst. Statist. Math. 31, 145–155. doi:10.1007/BF02480272

Beylkin G, Monzon L (2010). Approximation by exponential sums revisited. Appl. Comput. Harmon. Anal. 28, 131–149. doi:10.1016/j.acha.2009.08.011

Veen A, Schoenberg FP (2008). Estimation of space-time branching process models in seismology using an EM-type algorithm. JASA 103, 614–624. doi:10.1198/016214508000000148

Kirchner M (2017). An estimation procedure for the Hawkes process. Quant. Finance 17, 571–595. doi:10.1080/14697688.2016.1211312

Filimonov V, Sornette D (2015). Apparent criticality and calibration issues in the Hawkes self-excited point process model: application to high-frequency financial data. Quant. Finance 15(8), 1293–1314. doi:10.1080/14697688.2015.1032544

Examples

set.seed(1)
times <- sort(stats::runif(400, 0, 100))
fit <- chwkfit(times, 100, "exponential")
fit$braratio
#> [1] 0.3996747
fit$atbound # an interior maximum: the better of two starts
#> character(0)
# one start alone can stop with the decay on its bound, at a worse nll
chwkfit(times, 100, "exponential", start = c(log(2.4), 0.4, 4))$atbound
#> [1] "beta"
fit$nll
#> [1] -156.4648
chwkfit(times, 100, "lomax", meth = "soe", eps = 1e-8)$nll
#> [1] -156.388
chwkfit(times, 100, "exponential", base = "sinusoidal")$aic
#> [1] -305.1927

# the fast approximate INAR estimator
chwkfit(times, 100, meth = "inar")$braratio
#> [1] 0.005597589

# daily dates: spread the ties across each day first
days <- floor(times)
jit <- chwkjit(days, hor = 100)
chwkfit(jit, 100)$converged
#> [1] TRUE

# too few events, or times past the horizon, are refused
try(chwkfit(times[1:20], 100))
#> Error : too few events (20) for the fit
try(chwkfit(times, 50))
#> Error : `times` must lie within [0, hor] (they span 1.30776 to 99.6077); chwkjit(times, hor = ) keeps them inside