Current Models
Juniper L. Simonis, Hao Ye, Ethan P. White, and S. K. Morgan Ernest
21 August, 2024
Source:vignettes/current_models.Rmd
current_models.RmdModels
We currently analyze and forecast rodent data at Portal using fifteen models: AutoARIMA, Seasonal AutoARIMA, ESSS, NaiveARIMA, Seasonal NaiveARIMA, nbGARCH, nbsGARCH, pGARCH, psGARCH, pevGARCH, Random Walk, Logistic, Logistic Covariates, Logistic Competition, Logistic Competition Covariates (WeecologyLab 2019).
AutoARIMA
AutoARIMA (Automatic Auto-Regressive Integrated Moving Average) is a
flexible Auto-Regressive Integrated Moving Average (ARIMA) model. ARIMA
models are defined according to three model structure parameters – the
number of autoregressive terms (p), the degree of differencing (d), and
the order of the moving average (q), and are represented as ARIMA(p, d,
q) (Box and Jenkins 1970). All submodels
are fit and the final model is selected using the
auto.arima function in the forecast
package (Hyndman and Athanasopoulos 2013; Hyndman
2017). AutoARIMA is fit flexibly, such that the model parameters
can vary from fit to fit. Model forecasts are generated using the
forecast function from the forecast
package (Hyndman and Athanasopoulos 2013; Hyndman
2017). While the auto.arima function allows for
seasonal models, the seasonality is hard-coded to be on the same period
as the sampling, which is not the case for the Portal rodent surveys. As
a result, no seasonal models are evaluated within this model set, but
see the sAutoARIMA model. The model is fit using a normal distribution,
and as a result can generate negative-valued predictions and
forecasts.
Seasonal AutoARIMA
sAutoARIMA (Seasonal Automatic Auto-Regressive Integrated Moving
Average) is a flexible Auto-Regressive Integrated Moving Average (ARIMA)
model. ARIMA models are defined according to three model structure
parameters – the number of autoregressive terms (p), the degree of
differencing (d), and the order of the moving average (q), and are
represented as ARIMA(p, d, q) (Box and Jenkins
1970). All submodels are fit and the final model is selected
using the auto.arima function in the
forecast package (Hyndman and
Athanasopoulos 2013; Hyndman 2017). sAutoARIMA is fit flexibly,
such that the model parameters can vary from fit to fit. Model forecasts
are generated using the forecast function from the
forecast package (Hyndman and
Athanasopoulos 2013; Hyndman 2017). While the
auto.arima function allows for seasonal models, the
seasonality is hard-coded to be on the same period as the sampling,
which is not the case for the Portal rodent surveys. We therefore use
two Fourier series terms (sin and cos of the fraction of the year) as
external regressors to achieve the seasonal dynamics in the sAutoARIMA
model. The model is fit using a normal distribution, and as a result can
generate negative-valued predictions and forecasts.
ESSS
ESSS (Exponential Smoothing State Space) is a flexible exponential
smoothing state space model (Hyndman et al.
2008). The model is fit using the ets function in
the forecast package (Hyndman
2017) with the allow.multiplicative.trend argument
set to TRUE. Model forecasts are generated using the
forecast function from the forecast
package (Hyndman and Athanasopoulos 2013; Hyndman
2017). Models fit using ets implement what is known
as the “innovations” approach to state space modeling, which assumes a
single source of noise that is equivalent for the process and
observation errors (Hyndman et al. 2008).
In general, ESSS models are defined according to three model structure
parameters – error type, trend type, and seasonality type (Hyndman et al. 2008). Each of the parameters
can be an N (none), A (additive), or M (multiplicative) state (Hyndman et al. 2008). However, because of the
difference in period between seasonality and sampling of the Portal
rodents combined with the hard-coded single period of the
ets function, we do not include the seasonal components to
the ESSS model. ESSS is fit flexibly, such that the model parameters can
vary from fit to fit. The model is fit using a normal distribution, and
as a result can generate negative-valued predictions and forecasts.
NaiveARIMA
NaiveARIMA (Naive Auto-Regressive Integrated Moving Average) is a
fixed Auto-Regressive Integrated Moving Average (ARIMA) model of order
(0,1,0). The model is fit using the Arima function in the
forecast package (Hyndman and
Athanasopoulos 2013; Hyndman 2017). Model forecasts are generated
using the forecast function from the
forecast package (Hyndman and
Athanasopoulos 2013; Hyndman 2017). While the Arima
function allows for seasonal models, the seasonality is hard-coded to be
on the same period as the sampling, which is not the case for the Portal
rodent surveys. As a result, NaiveARIMA does not include seasonal terms,
but see the sNaiveARIMA model. The model is fit using a normal
distribution, and as a result can generate negative-valued predictions
and forecasts.
Seasonal NaiveARIMA
sNaiveARIMA (Seasonal Naive Auto-Regressive Integrated Moving
Average) is a fixed Auto-Regressive Integrated Moving Average (ARIMA)
model of order (0,1,0). The model is fit using the Arima
function in the forecast package (Hyndman and Athanasopoulos 2013; Hyndman 2017).
Model forecasts are generated using the forecast function
from the forecast package (Hyndman and Athanasopoulos 2013; Hyndman 2017).
While the Arima function allows for seasonal models, the
seasonality is hard-coded to be on the same period as the sampling,
which is not the case for the Portal rodent surveys. We therefore use
two Fourier series terms (sin and cos of the fraction of the year) as
external regressors to achieve the seasonal dynamics in the NaiveARIMA
model. The model is fit using a normal distribution, and as a result can
generate negative-valued predictions and forecasts.
nbGARCH
nbGARCH (Negative Binomial Auto-Regressive Conditional
Heteroskedasticity) is a generalized autoregressive conditional
heteroskedasticity (GARCH) model with overdispersion (i.e., a
negative binomial response). The model is fit to the interpolated data
using the tsglm function in the tscount
package (Liboschik et al. 2017). GARCH
models are generalized ARMA models and are defined according to their
link function, response distribution, and two model structure parameters
– the number of autoregressive terms (p) and the order of the moving
average (q), and are represented as GARCH(p, q) (Liboschik et al. 2017). The nbGARCH model is
fit using the log link and a negative binomial response (modeled as an
over-dispersed Poisson), as well as with p = 1 (first-order
autoregression) and q = 13 (approximately yearly moving average). The
tsglm function in the tscount package
(Liboschik et al. 2017) uses a
(conditional) quasi-likelihood based approach to inference and models
the overdispersion as an additional parameter in a two-step approach.
This two-stage approach has only been minimally evaluated, although
preliminary simulation-based studies are promising (Liboschik, Fokianos, and Fried 2017). Forecasts
are made using the portalcasting method function
forecast.tsglm as forecast.
nbsGARCH
nbsGARCH (Negative Binomial Seasonal Auto-Regressive Conditional
Heteroskedasticity) is a generalized autoregressive conditional
heteroskedasticity (GARCH) model with overdispersion (i.e., a
negative binomial response) and seasonal predictors modeled using two
Fourier series terms (sin and cos of the fraction of the year) fit to
the interpolated data. The model is fit using the tsglm
function in the tscount package (Liboschik et al. 2017). GARCH models are
generalized ARMA models and are defined according to their link
function, response distribution, and two model structure parameters –
the number of autoregressive terms (p) and the order of the moving
average (q), and are represented as GARCH(p, q) (Liboschik et al. 2017). The nbsGARCH model is
fit using the log link and a negative binomial response (modeled as an
over-dispersed Poisson), as well as with p = 1 (first-order
autoregression) and q = 13 (approximately yearly moving average). The
tsglm function in the tscount package
(Liboschik et al. 2017) uses a
(conditional) quasi-likelihood based approach to inference and models
the overdispersion as an additional parameter in a two-step approach.
This two-stage approach has only been minimally evaluated, although
preliminary simulation-based studies are promising (Liboschik, Fokianos, and Fried 2017). Forecasts
are made using the portalcasting method function
forecast.tsglm as forecast.
pGARCH
pGARCH (Poisson Auto-Regressive Conditional Heteroskedasticity) is a
generalized autoregressive conditional heteroskedasticity (GARCH) model.
The model is fit using the tsglm function in the
tscount package (Liboschik et
al. 2017). GARCH models are generalized ARMA models and are
defined according to their link function, response distribution, and two
model structure parameters – the number of autoregressive terms (p) and
the order of the moving average (q), and are represented as GARCH(p, q)
(Liboschik et al. 2017). The pGARCH model
is fit using the log link and a Poisson response, as well as with p = 1
(first-order autoregression) and q = 13 (approximately yearly moving
average). Forecasts are made using the portalcasting
method function forecast.tsglm as
forecast.
psGARCH
psGARCH (Poisson Seasonal Auto-Regressive Conditional
Heteroskedasticity) is a generalized autoregressive conditional
heteroskedasticity (GARCH) model with seasonal predictors modeled using
two Fourier series terms (sin and cos of the fraction of the year) fit
to the interpolated data. The model is fit using the tsglm
function in the tscount package (Liboschik et al. 2017). GARCH models are
generalized ARMA models and are defined according to their link
function, response distribution, and two model structure parameters –
the number of autoregressive terms (p) and the order of the moving
average (q), and are represented as GARCH(p, q) (Liboschik et al. 2017). The pGARCH model is fit
using the log link and a Poisson response, as well as with p = 1
(first-order autoregression) and q = 13 (approximately yearly moving
average). Forecasts are made using the portalcasting
method function forecast.tsglm as
forecast.
pevGARCH
pevGARCH (Poisson Environmental Variable Auto-Regressive Conditional
Heteroskedasticity) is a generalized autoregressive conditional
heteroskedasticity (GARCH) model. The response variable is Poisson, and
a variety of environmental variables are considered as covariates. The
overall model is fit using the portalcasting function
meta_tsglm that iterates over the submodels, which are fit
using the tsglm function in the tscount
package (Liboschik et al. 2017). GARCH
models are generalized ARMA models and are defined according to their
link function, response distribution, and two model structure parameters
– the number of autoregressive terms (p) and the order of the moving
average (q), and are represented as GARCH(p, q) (Liboschik et al. 2017). The pevGARCH model is
fit using the log link and a Poisson response, as well as with p = 1
(first-order autoregression) and q = 13 (yearly moving average). The
environmental variables potentially included in the model are min, mean,
and max temperatures, precipitation, and NDVI. The tsglm
function in the tscount package (Liboschik et al. 2017) uses a (conditional)
quasi-likelihood based approach to inference. This approach has only
been minimally evaluated for models with covariates, although
preliminary simulation-based studies are promising (Liboschik, Fokianos, and Fried 2017). The
overall model is composed of 11 submodels from a (nonexhaustive)
combination of the environmental covariates – [1] max temp, mean temp,
precipitation, NDVI; [2] max temp, min temp, precipitation, NDVI; [3]
max temp, mean temp, min temp, precipitation; [4] precipitation, NDVI;
[5] min temp, NDVI; [6] min temp; [7] max temp; [8] mean temp; [9]
precipitation; [10] NDVI; and [11] -none- (intercept-only). The single
best model of the 11 is selected based on AIC. Forecasts are made using
the portalcasting method function
forecast.tsglm as forecast.
Random Walk
Random Walk fits a hierarchical log-scale density random walk model
with a Poisson observation process using the JAGS (Just Another Gibbs
Sampler) infrastructure (Plummer 2003).
Similar to the NaiveArima model, Random Walk has an ARIMA order of
(0,1,0), but in Random Walk, it is the underlying density that takes a
random walk on the log scale, whereas in NaiveArima, it is the raw
counts that take a random walk on the observation scale. There are two
process parameters – mu (the density of the species at the beginning of
the time series) and sigma (the standard deviation of the random walk,
which is Gaussian on the log scale). The observation model has no
additional parameters. The prior distributions for mu is informed by the
available data collected prior to the start of the data used in the time
series. mu is normally distributed with a mean equal to the average
log-scale density and a standard deviation of 0.04. sigma was given a
uniform distribution between 0 and 1. The Random Walk model is fit and
forecast using the portalcasting functions
fit_runjags and forecast.runjags (called as
forecast), respectively.
Logistic
Logistic fits a hierarchical log-scale density logistic growth model
with a Poisson observation process using the JAGS (Just Another Gibbs
Sampler) infrastructure (Plummer 2003).
Building upon the Random Walk model, Logistic expands upon the “process
model” underlying the Poisson observations. There are four process
parameters – mu (the density of the species at the beginning of the time
series), sigma (the standard deviation of the random walk, which is
Gaussian on the log scale), r (growth rate), and K (carrying capacity).
The observation model has no additional parameters. The prior
distributions for mu and K are slightly informed in that they are vague
but centered using the available data and sigma and r are set with vague
priors. mu is normally distributed with a mean equal to the average
log-scale density and a standard deviation of 0.04. K is modeled on the
log-scale with a prior mean equal to the log maximum count and a
standard deviation of 0.04. r is given a normal prior with mean 0 and
standard deviation 0.04. sigma was given a uniform distribution between
0 and 1. The Logistic model is fit and forecast using the
portalcasting functions fit_runjags and
forecast.runjags (called as forecast),
respectively.
Logistic Covariates
Logistic Covariates fits a hierarchical log-scale density logistic
growth model with a Poisson observation process using the JAGS (Just
Another Gibbs Sampler) infrastructure (Plummer
2003). Building upon the Logistic model, Logistic Covariates
expands upon the “process model” underlying the Poisson observations.
There are six process parameters – mu (the density of the species at the
beginning of the time series), sigma (the standard deviation of the
random walk, which is Gaussian on the log scale), and then intercept and
slope parameters for r (growth rate) and K (carrying capacity) as a
function of covariates (r being a function of the integrated warm rain
over the past 3 lunar months and K being a function of average NDVI over
the 13 lunar months). The observation model has no additional
parameters. The prior distributions for mu and the K interceptare
slightly informed in that they are vague but centered using the
available data and sigma and r are set with vague priors. mu is normally
distributed with a mean equal to the average log-scale density and
standard deviation 0.04. The K intercept is modeled on the log-scale
with a prior mean equal to the maximum count and standard deviation
0.04. The r intercept is given a normal prior with mean 0 and standard
deviation 0.04. sigma was given a uniform distribution between 0 and 1.
The slopes for r and log-scale K were given priors with mean 0 and
standard deviation 1. The Logistic Covariates model is fit and forecast
using the portalcasting functions
fit_runjags and forecast.runjags (called as
forecast), respectively.
Logistic Competition
Logistic Competition fits a hierarchical log-scale density logistic
growth model with a Poisson observation process using the JAGS (Just
Another Gibbs Sampler) infrastructure (Plummer
2003). Building upon the Logistic model, Logistic Competition
expands upon the “process model” underlying the Poisson observations.
There are six process parameters – mu (the density of the species at the
beginning of the time series), sigma (the standard deviation of the
random walk, which is Gaussian on the log scale), and then intercept and
slope parameters for r (growth rate) and K (carrying capacity) as a
function of competitior density (K being a function of current DO
counts). The observation model has no additional parameters. The prior
distributions for mu and the K intercept are slightly informed in that
they are vague but centered using the available data and sigma and r are
set with vague priors. mu is normally distributed with a mean equal to
the average log-scale density and standard deviation 0.04. The K
intercept is modeled on the log-scale with a prior mean equal to the
maximum of past counts and standard deviation 0.04. The r
intercept is given a normal prior with mean 0 and standard deviation
0.04. sigma was given a uniform distribution between 0 and 1. The slope
for log-scale K was given a prior with mean 0 and standard deviation 1.
The Logistic Competition model is fit and forecast using the
portalcasting functions fit_runjags and
forecast.runjags (called as forecast),
respectively.
Logistic Competition Covariates
Logistic Competition Covariates fits a hierarchical log-scale density
logistic growth model with a Poisson observation process using the JAGS
(Just Another Gibbs Sampler) infrastructure (Plummer 2003). Building upon the Logistic
model, Logistic Competition Covariates expands upon the “process model”
underlying the Poisson observations. There are seven process parameters
– mu (the density of the species at the beginning of the time series),
sigma (the standard deviation of the random walk, which is Gaussian on
the log scale), and then intercept and slope parameters for r (growth
rate) and K (carrying capacity) with r being a function of the
integrated warm rain over the past 3 lunar months and K being a function
of average NDVI over the past 13 lunar months as well as a function of
current DO count). The observation model has no additional parameters.
The prior distributions for mu and the K intercept are are slightly
informed in that they are vague but centered using the available data
and sigma and r are set with vague priors. mu is normally distributed
with a mean equal to the average log-scale density and standard
deviation 0.04. The K intercept is modeled on the log-scale with a prior
mean equal to the maximum count and standard deviation 0.04. The r
intercept is given a normal prior with mean 0 and standard deviation
0.04. Sigma was given a uniform distribution between 0 and 0.001. The
slope for r was given a normal prior with mean 0 and standard deviation
1, whereas the normal priors for the log K slopes were given mean 0 and
standard deviation 0.0625. The Logistic Competition Covariates model is
fit and forecast using the portalcasting functions
fit_runjags and forecast.runjags (called as
forecast), respectively.
Ensemble
In addition to the base models, we include a starting-point ensemble. Prior to v0.9.0, the ensemble was based on AIC weights, but in the shift to separating the interpolated from non-interpolated data in model fitting, we had to transfer to an unweighted average ensemble model. The ensemble mean is calculated as the mean of all model means and the ensemble variance is estimated as the sum of the mean of all model variances and the variance of the estimated mean, calculated using the unbiased estimate of sample variances. Given that the current ensemble is unweighted and includes a number of very naive models, we do not currently consider it the best model for forecasting.