R: MPIN model estimation via an ECM algorithm

mpin_ecm {PINstimation}

R Documentation

MPIN model estimation via an ECM algorithm

Description

Estimates the multilayer probability of informed trading (MPIN) using an Expectation Conditional Maximization algorithm, as in Ghachem and Ersan (2022a).

Usage

mpin_ecm(data, layers = NULL, xtraclusters = 4, initialsets = NULL,
                      ..., verbose = TRUE)

Arguments

`data`	A dataframe with 2 variables: the first corresponds to buyer-initiated trades (buys), and the second corresponds to seller-initiated trades (sells).
`layers`	An integer referring to the assumed number of information layers in the data. If the argument `layers` is given, then the ECM algorithm will use the number of layers provided. If `layers` is omitted, the function `mpin_ecm()` will simultaneously optimize the number of layers as well as the parameters of the `MPIN` model.
`xtraclusters`	An integer used to divide trading days into `#(1 + layers + xtraclusters)` clusters, thereby resulting in `#comb((layers + xtraclusters, layers)` initial parameter sets in line with Ersan and Alici (2016), and Ersan (2016). The default value is `4` as chosen in Ersan (2016).
`initialsets`	A dataframe containing initial parameter sets for estimation of the `MPIN` model. The default value is `NULL`. If `initialsets` is `NULL`, the initial parameter sets are provided by the function `initials_mpin()`.
`...`	Additional arguments passed on to the function `mpin_ecm`. The recognized arguments are `hyperparams`, and `is_parallel`. `hyperparams` is a list containing the hyperparameters of the ECM algorithm. When not empty, it contains one or more of the following elements: `minalpha`, `maxeval`, `tolerance`, `criterion`, and `maxlayers`. More about these elements are in the details section. `is_parallel` is a logical variable that specifies whether the computation is performed using parallel or sequential processing. The default value is `FALSE`. For more details, please refer to the vignette 'Parallel processing' in the package, or online.
`verbose`	(`logical`) a binary variable that determines whether detailed information about the steps of the estimation of the MPIN model is displayed. No output is produced when `verbose` is set to `FALSE`. The default value is `TRUE`.

Details

The argument 'data' should be a numeric dataframe, and contain at least two variables. Only the first two variables will be considered: The first variable is assumed to correspond to the total number of buyer-initiated trades, while the second variable is assumed to correspond to the total number of seller-initiated trades. Each row or observation correspond to a trading day. NA values will be ignored.

The initial parameters for the expectation-conditional maximization algorithm are computed using the function initials_mpin() with default settings. The factorization of the MPIN likelihood function used is developed by Ersan (2016), and is implemented in fact_mpin().

The argument hyperparams contains the hyperparameters of the ECM algorithm. It is either empty or contains one or more of the following elements:

minalpha (numeric) It stands for the minimum share of days belonging to a given layer, i.e., layers falling below this threshold are removed during the iteration, and the model is estimated with a lower number of layers. When missing, minalpha takes the default value of 0.001.
maxeval: (integer) It stands for maximum number of iterations of the ECM algorithm for each initial parameter set. When missing, maxeval takes the default value of 100.
tolerance (numeric) The ECM algorithm is stopped when the (relative) change of log-likelihood is smaller than tolerance. When missing, tolerance takes the default value of 0.001.
criterion (character) It is the model selection criterion used to find the optimal estimate for the MPIN model. It take one of these values "BIC", "AIC" and "AWE"; which stand for Bayesian Information Criterion, Akaike Information Criterion and Approximate Weight of Evidence, respectively (Akogul and Erisoglu 2016). When missing, criterion takes the default value of "BIC".
maxlayers (integer) It is the upper limit of number of layers used for estimation in the ECM algorithm. If the argument layers is missing, the ECM algorithm will estimate MPIN models for all layers in the integer set from 1 to maxlayers. When missing, maxlayers takes the default value of 8.
maxinit (integer) It is the maximum number of initial sets used for each individual estimation in the ECM algorithm. When missing, maxinit takes the default value of 100.

If the argument layers is given, then the Expectation Conditional Maximization algorithm will use the number of layers provided. If layers is omitted, the function mpin_ecm() will simultaneously optimize the number of layers as well as the parameters of the MPIN model. Practically, the function mpin_ecm() uses the ECM algorithm to optimize the MPIN model parameters for each number of layers within the integer set from 1 to 8 (or to maxlayers if specified in the argument hyperparams); and returns the optimal model with the lowest Bayesian information criterion (BIC) (or the lowest information criterion criterion if specified in the argument hyperparams).

Value

Returns an object of class estimate.mpin.ecm.

References

Akogul S, Erisoglu M (2016). “A comparison of information criteria in clustering based on mixture of multivariate normal distributions.” Mathematical and Computational Applications, 21(3), 34.

Ersan O (2016). “Multilayer Probability of Informed Trading.” Available at SSRN 2874420.

Ersan O, Alici A (2016). “An unbiased computation methodology for estimating the probability of informed trading (PIN).” Journal of International Financial Markets, Institutions and Money, 43, 74–94. ISSN 10424431.

Ghachem M, Ersan O (2022a). “Estimation of the probability of informed trading models via an expectation-conditional maximization algorithm.” Available at SSRN 4117952.

Examples

# There is a preloaded quarterly dataset called 'dailytrades' with 60
# observations. Each observation corresponds to a day and contains the
# total number of buyer-initiated trades ('B') and seller-initiated
# trades ('S') on that day. To know more, type ?dailytrades

xdata <- dailytrades

# Estimate the MPIN model using the expectation-conditional maximization
# (ECM) algorithm.

# ------------------------------------------------------------------------ #
# Estimate the MPIN model, assuming that there exists 2 information layers #
# in the dataset                                                           #
# ------------------------------------------------------------------------ #

estimate <- mpin_ecm(xdata, layers = 2, verbose = FALSE)

# Show the estimation output

show(estimate)

# Display the optimal parameters from the Expectation Conditional
# Maximization algorithm

show(estimate@parameters)

# Display the global multilayer probability of informed trading

show(estimate@mpin)

# Display the multilayer probability of informed trading per layer

show(estimate@mpinJ)

# Display the first five rows of the initial parameter sets used in the
# expectation-conditional maximization estimation

show(round(head(estimate@initialsets, 5), 4))

# ------------------------------------------------------------------------ #
# Omit the argument 'layers', so the ECM algorithm optimizes both the      #
# number of layers and the MPIN model parameters.                          #
# ------------------------------------------------------------------------ #

estimate <- mpin_ecm(xdata, verbose = FALSE)

# Show the estimation output

show(estimate)

# Display the optimal parameters from the estimation of the MPIN model using
# the expectation-conditional maximization (ECM) algorithm

show(estimate@parameters)

# Display the multilayer probability of informed trading

show(estimate@mpin)

# Display the multilayer probability of informed trading per layer

show(estimate@mpinJ)

# Display the first five rows of the initial parameter sets used in the
# expectation-conditional maximization estimation.

show(round(head(estimate@initialsets, 5), 4))

# ------------------------------------------------------------------------ #
# Tweak in the hyperparameters of the ECM algorithm                        #
# ------------------------------------------------------------------------ #

# Create a variable ecm.params containing the hyperparameters of the ECM
# algorithm. This will surely make the ECM algorithm take more time to give
# results

ecm.params <- list(tolerance = 0.0000001)

# If we suspect that the data contains more than eight information layers, we
# can raise the number of models to be estimated to 10 as an example, i.e.,
# maxlayers = 10.

ecm.params$maxlayers <- 10

# We can also choose Approximate Weight of Evidence (AWE) for model
# selection instead of the default Bayesian Information Criterion (BIC)

ecm.params$criterion <- 'AWE'

# We can also increase the maximum number of initial sets to 200, in
# order to obtain higher level of accuracy for models with high number of
# layers.  We set the sub-argument 'maxinit' to `200`. Remember that its
# default value is `100`.

ecm.params$maxinit <- 200

estimate <- mpin_ecm(xdata, xtraclusters = 2, hyperparams = ecm.params,
                                                      verbose = FALSE)

# We can change the model selection criterion by calling selectModel()

estimate <- selectModel(estimate, "AIC")

# We get the mpin_ecm estimation results for the MPIN model with 2 layers
# using the slot models. We then show the first five rows of the
# corresponding slot details.

models <- estimate@models
show(round(head(models[[2]]@details, 5), 4))

# We can also use the function getSummary to get an idea about the change in
# the estimation parameters as a function of the number of layers in the
# MPIN model. The function getSummary returns a dataframe that contains,
# among others, the number of layers of the model, the number of layers in
# the optimal model,the MPIN value, and the values of the different
# information criteria, namely AIC, BIC and AWE.

summary <- getSummary(estimate)

# We can plot the MPIN value and the layers at the optimal model as a
# function of the number of layers to see whether additional layers in the
# model actually contribute to a better precision in the probability of
# informed trading. Remember that the hyperparameter 'minalpha' is
# responsible for dropping layers with "frequency" lower than 'minalpha'.

plot(summary$layers, summary$MPIN,
   type = "o", col = "red",
   xlab = "MPIN model layers", ylab = "MPIN value"
 )

plot(summary$layers, summary$em.layers,
   type = "o", col = "blue",
   xlab = "MPIN model layers", ylab = "layers at the optimal model"
)

[Package PINstimation version 0.1.2 Index]