NEWS (original) (raw)

newcustomer() prediction: Include the initial transaction in the predicted number of orders

clvdata(data.end): Add parameter data.endto specify a data end beyond the last actual transaction
summary(): Always set zval andpval to NA for the main model parameters
hessian(): Add method to calculate hessian matrix for already fitted models
Add 3 new vignettes covering: Advanced modelling techniques, model intuition, and the internal class system

Fix CRAN notes: Replace arma::is_finite() ->std::isfinite()
Dyncov PNBD: Rename predicted.CLV ->predicted.period.CLV
predict(): Rename{predicted, actual}.total.spending ->{predicted, actual}.period.spending

newcustomer.spending(): Predict average spending per transaction for customers without order history
Improved optimizer defaults (higher iteration count) for PNBD dyncov

Updated the apparel example data
Prediction bootstrapping: Calculate confidence intervals using regular rather than “reversed-quantiles”

More memory efficient and faster creation of repeat transactions inclv.data
Use existing repeat transactions when calling gg withremove.first.transaction = TRUE
Simplify the formula interfaces latentAttrition() andspending()
Add predicted.total.spending to predictions
Harmonize parameter names used in various S3 methods
Bootstrapping: Add facilities to estimate parameter uncertainty for all models
Ability to predict future transactions of customers with no existing transaction history
New start parameters for all latent attrition models
Pareto/NBD dyncov: Improved numeric stability of PAlive
GGomNBD: Implement erratum by Jost Adler to predict CET correctly
GGomNBD: Improve numerical stability and runtime of LL integral
GGomNBD: Implement PMF as derived by Jost Adler
lrtest(): Likelihood ratio testing for latent attrition models
Accept data.table::IDate as data inputs toclvdata
summary.clv.data:Much faster by improving the calculation of the mean inter-purchase time
Reduced fitting times for all models by using a compressed CBS as input to the LL sum
Faster hessian calculation if a model was using correlation

Estimating the Pareto/NBD dyncov with correlation was not possible
GGomNBD: Free workspace after it is not used anymore to avoid memory-leak
SetDynamicCovariates: Verify there is no covariate data for nonexistent customers

We add an interface to specify models using a formula notation (latentAttrition() and spending())
New method to plot customer’s transaction timings (plot.clv.data(which='timings'))
Draw diagnostic plots of multiple models in single plot (plot(other.models=list(), label=c()))
MUCH faster fitting for the Pareto/NBD with time-varying covariates because we implemented the LL in Rcpp

Three new diagnostic plots for transaction data to analyse frequency, spending and interpurchase time
New diagnostic plot for fitted transaction models (PMF plot)
New function to calculate the probability mass function of selected models
Calculate summary statistics only for the transaction data of selected customers
Canonical transformation from data.frame/data.table to transaction data object and vice-versa
Canonical subset for the data stored in the transaction data object
Pareto/NBD DERT: Improved numerical stability

Partially refactor the LL of the extended Pareto/NBD in Rcpp with code kindly donated by Elliot Shin Oblander
Improved documentation

Optimization methods nlm and nlminb can now be used. Thanks to Elliot Shin Oblander for reporting

Refactor the Gamma-Gamma (GG) model to predict mean spending per transaction into an independent model
The prediction for transaction models can now be combined with separately fit spending models
Write the unconditional expectation functions in Rcpp for faster plotting (Pareto/NBD and Beta-Geometric/NBD)
Improved documentation and walkthrough

Pareto/NBD log-likelihood: For the case Tcal = t.x and for the case alpha == beta
Static or dynamic covariates with syntactically invalid names (spaces, start with numbers, etc) could not be fit

Beta-Geometric/NBD (BG/NBD) model to predict repeat transactions without and with static covariates
Gamma-Gompertz (GGompertz) model to predict repeat transactions without and with static covariates
Predictions are now possible for all periods >= 0 whereas before a minimum of 2 periods was required
Initial release of the CLVTools package

Pareto/NBD model to predict repeat transactions without and with static or dynamic covariates
Gamma-Gamma model to predict average spending
Predicting CLV and future transactions per customer
Data class to preprocess transaction data and to provide summary statistics
Plot of expected repeat transactions as by the fitted model compared against actuals