In this vignette we show a pipeline for predicting individual claim counts using the ReSurv package.

This vignette was created for the 2024 Reserving Call Paper Program on Technology and the Reserving Actuary of the Casualty Actuarial Society (CAS) and the CAS Reserves Working Group. We would like to thank the team of reviewers that provided feedback that improved our package.

Installation

To install the developer version of the ReSurv package.

library(devtools)
devtools::install_github("edhofman/resurv")
library(ReSurv)
packageVersion("ReSurv")

To handle the Python dependencies for the first package installation, we suggest to create a dedicated virtual environment. The remaining part of this Section can be disregard from users that are not interested in using our models based on Neural Networks.

ReSurv::install_pyresurv()

The default name of the virtual environment is "pyresurv".

We then suggest to refresh the R session and to import the ReSurv package in R using

library(ReSurv)
reticulate::use_virtualenv("pyresurv")

Managing Multiple Package Dependencies

For the case of multiple packages using isolated-package-environments we suggest the following procedure. Below an example for pysparklyr.

envname <- "./venv"
ReSurv::install_pyresurv(envname = envname)
pysparklyr::install_pyspark(envname = envname)

Data Simulation

We simulate a synthetic data set from scenario Alpha.

input_data <- data_generator(random_seed = 7,
                             scenario = "alpha",
                             time_unit = 1 / 360,
                             years = 4,
                             period_exposure = 200)

We inspect the data set.

str(input_data)

Data Pre-Processing

The data is pre-processed using the IndividualDataPP function.

individual_data <- IndividualDataPP(data = input_data,
                                    categorical_features = "claim_type",
                                    continuous_features = "AP",
                                    accident_period = "AP",
                                    calendar_period = "RP",
                                    input_time_granularity = "days",
                                    output_time_granularity = "quarters",
                                    years = 4)

Selection of the hyper-parameters

This Sections first shows the usage of the ReSurvCV function for cross-validation. Secondly, we show how to combine the ReSurvCV function into the ParBayesianOptimizationpackage routine for performing the Bayesian Optimization of Hyperparameters.

XGB: K-Fold cross-validation (CV)

Example of K-Fold cross-validation (CV) for the XGB model.

resurv_cv_xgboost <- ReSurvCV(IndividualDataPP = individual_data,
                              model = "XGB",
                              hparameters_grid = list(booster = "gbtree",
                                                      eta = c(.001),
                                                      max_depth = c(3),
                                                      subsample = c(1),
                                                      alpha = c(0),
                                                      lambda = c(0),
                                                      min_child_weight = c(.5)),
                              print_every_n = 1L,
                              nrounds = 1,
                              verbose = FALSE,
                              verbose.cv = TRUE,
                              early_stopping_rounds = 1,
                              folds = 5,
                              parallel = TRUE,
                              ncores = 2,
                              random_seed = 1)

NN: K-Fold cross-validation

Example of K-Fold cross-validation (CV) for the NN model.

resurv_cv_nn <- ReSurvCV(IndividualDataPP = individual_data,
                         model = "NN",
                         hparameters_grid = list(num_layers = c(1, 2),
                                                 num_nodes = c(2, 4),
                                                 optim = "Adam",
                                                 activation = "ReLU",
                                                 lr = .5,
                                                 xi = .5,
                                                 eps = .5,
                                                 tie = "Efron",
                                                 batch_size = as.integer(5000),
                                                 early_stopping = "TRUE",
                                                 patience = 20),
                         epochs = as.integer(300),
                         num_workers = 0,
                         verbose = FALSE,
                         verbose.cv = TRUE,
                         folds = 3,
                         parallel = FALSE,
                         random_seed = as.integer(Sys.time()))

ReSurv and Bayesian Parameters Optimisation

XGB: Bayesian Parameters Optimisation

Example of Bayesian Optimization of Hyperparameters for the XGB model.

bounds <- list(eta = c(0, 1),
               max_depth = c(1L, 25L),
               min_child_weight = c(0, 50),
               subsample = c(0.51, 1),
               lambda = c(0, 15),
               alpha = c(0, 15))


obj_func <- function(eta,
                     max_depth,
                     min_child_weight,
                     subsample,
                     lambda,
                     alpha) {

  xgbcv <- ReSurvCV(
    IndividualDataPP = individual_data,
    model = "XGB",
    hparameters_grid = list(
      booster = "gbtree",
      eta = eta,
      max_depth = max_depth,
      subsample = subsample,
      alpha = lambda,
      lambda = alpha,
      min_child_weight = min_child_weight
    ),
    print_every_n = 1L,
    nrounds = 500,
    verbose = FALSE,
    verbose.cv = TRUE,
    early_stopping_rounds = 30,
    folds = 3,
    parallel = FALSE,
    random_seed = as.integer(Sys.time())
  )

  lst <- list(Score = -xgbcv$out.cv.best.oos$test.lkh,
              train.lkh = xgbcv$out.cv.best.oos$train.lkh)

  return(lst)
}



bayes_out <- bayesOpt(FUN = obj_func,
                      bounds = bounds,
                      initPoints = 50,
                      iters.n = 1000,
                      iters.k = 50,
                      otherHalting = list(timeLimit = 18000))

NN: Bayesian Parameters Optimisation

Example of Bayesian Optimization of Hyperparameters for the XGB model.

bounds <- list(num_layers = c(2L, 10L),
               num_nodes = c(2L, 10L),
               optim = c(1L, 2L),
               activation = c(1L, 2L),
               lr = c(0.005, 0.5),
               xi = c(0, 0.5),
               eps = c(0, 0.5))


obj_func <- function(num_layers,
                     num_nodes,
                     optim,
                     activation,
                     lr,
                     xi,
                     eps) {

  optim <- switch(optim,
                  "Adam",
                  "SGD")

  activation <- switch(activation, "LeakyReLU", "SELU")
  batch_size <- as.integer(5000L)
  number_layers <- as.integer(num_layers)
  num_nodes <- as.integer(num_nodes)

  deepsurv_cv <- ReSurvCV(IndividualData = individual_data,
                          model = "NN",
                          hparameters_grid = list(num_layers = number_layers,
                                                  num_nodes = num_nodes,
                                                  optim = optim,
                                                  activation = activation,
                                                  lr = lr,
                                                  xi = xi,
                                                  eps = eps,
                                                  tie = "Efron",
                                                  batch_size = batch_size,
                                                  early_stopping = "TRUE",
                                                  patience  = 20),
                          epochs = as.integer(300),
                          num_workers = 0,
                          verbose = FALSE,
                          verbose.cv = TRUE,
                          folds = 3,
                          parallel = FALSE,
                          random_seed = as.integer(Sys.time()))


  lst <- list(Score = -deepsurv_cv$out.cv.best.oos$test.lkh,
              train.lkh = deepsurv_cv$out.cv.best.oos$train.lkh)

  return(lst)
}

bayes_out <- bayesOpt(FUN = obj_func,
                      bounds = bounds,
                      initPoints = 50,
                      iters.n = 1000,
                      iters.k = 50,
                      otherHalting = list(timeLimit = 18000))

Estimation

In this Section we show the usage of the ReSurv method for the estimation of our models parameters. The hyperparameters for NN and XGB were derived with the Bayesian Optimisation procedure.

COX

resurv_fit_cox <- ReSurv(individual_data,
                     hazard_model = "COX")

XGB

hparameters_xgb <- list(
  params = list(
    booster = "gbtree",
    eta = 0.9611239,
    subsample = 0.62851,
    alpha = 5.836211,
    lambda = 15,
    min_child_weight = 29.18158,
    max_depth = 1
  ),
  print_every_n = 0,
  nrounds = 3000,
  verbose = FALSE,
  early_stopping_rounds = 500
)


resurv_fit_xgb <- ReSurv(individual_data,
                         hazard_model = "XGB",
                         hparameters = hparameters_xgb)

NN

hparameters_nn = list(num_layers= 2,
            early_stopping= TRUE,
            patience=350,
            verbose=FALSE,
            network_structure=NULL,
            num_nodes= 10,
            activation ="LeakyReLU",
            optim ="SGD",
            lr =0.02226655,
            xi=0.4678993,
            epsilon= 0,
            batch_size= 5000L,
            epochs= 5500L,
            num_workers= 0,
            tie="Efron" )


resurv_fit_nn <- ReSurv(individual_data,
                     hazard_model = "NN",
                     hparameters = hparameters_nn)

Prediction

In this Section we show how to use our models for predictions with the predict method and for different output time granularities (quarterly, yearly and monthly). The output can be displayed using the methods summary and print.

resurv_fit_predict_q <- predict(resurv_fit_cox)

individual_data_y <- IndividualDataPP(input_data,
                                      id = "claim_number",
                                      continuous_features = "AP_i",
                                      categorical_features = "claim_type",
                                      accident_period = "AP",
                                      calendar_period = "RP",
                                      input_time_granularity = "days",
                                      output_time_granularity = "years",
                                      years = 4,
                                      continuous_features_spline = NULL,
                                      calendar_period_extrapolation = FALSE)

resurv_fit_predict_y <- predict(resurv_fit_cox,
                                newdata = individual_data_y,
                                grouping_method = "probability")

individual_data_m <- IndividualDataPP(input_data,
                                      id = "claim_number",
                                      continuous_features = "AP_i",
                                      categorical_features = "claim_type",
                                      accident_period = "AP",
                                      calendar_period = "RP",
                                      input_time_granularity = "days",
                                      output_time_granularity = "months",
                                      years = 4,
                                      continuous_features_spline = NULL,
                                      calendar_period_extrapolation = FALSE)

resurv_fit_predict_m <- predict(resurv_fit_cox,
                                newdata = individual_data_m,
                                grouping_method = "probability")



model_s <- summary(resurv_fit_predict_y)
print(model_s)

Similar functions for the other models

resurv_predict_xgb <- predict(resurv_fit_xgb)
resurv_predict_nn <- predict(resurv_fit_nn)

Plot the predicted development factors

In this Section we show an example of development factors plot for the COX model.

resurv_fit_predict_q$long_triangle_format_out$output_granularity %>% 
    filter(AP_o==15 & claim_type == 1) %>% 
    filter(row_number()==1) %>%
    select(group_o)

plot(resurv_fit_predict_q, 
         granularity = "output",
         title_par = "COX: Accident Quarter 15 Claim Type 1",
         group_code = 30)

CRPS

In this Section we show the CRPS calculation using the survival_crps method.

crps <- survival_crps(resurv_fit_cox)
m_crps <- mean(crps$crps)
m_crps

Plot the development factors

We first extract the output group code for an arbitrary group.

resurv_fit_predict_q$long_triangle_format_out$output_granularity %>% 
    filter(AP_o==15 & claim_type == 1) %>% 
    filter(row_number()==1) %>%
    select(group_o)

We use the group code to plot the feature dependent development factors.

plot(resurv_fit_predict_q, 
         granularity = "output",
         title_par = "COX: Accident Quarter 15 Claim Type 1",
         group_code = 30)

Appendix

Code to replicate Figure 1

p1 <- input_data %>%
  as.data.frame() %>%
  mutate(claim_type = as.factor(claim_type)) %>%
  ggplot(aes(x = RT - AT, color = claim_type)) +
  stat_ecdf(size = 1) +
  labs(title = "", x = "Notification delay (in days)", y = "ECDF") +
  xlim(0.01, 1500) +
  scale_color_manual(
    values = c("royalblue", "#a71429"),
    labels = c("Claim type 0", "Claim type 1")
  ) +
  scale_linetype_manual(values = c(1, 3),
                        labels = c("Claim type 0", "Claim type 1")) +
  guides(
    color = guide_legend(
      title = "Claim type",
      override.aes = list(
        color = c("royalblue", "#a71429"),
        linewidth = 2
      )
    ),
    linetype = guide_legend(
      title = "Claim type",
      override.aes = list(linetype = c(1, 3), linewidth = 0.7)
    )
  ) +
  theme_bw() +
  theme(
    axis.text = element_text(size = 20),
    axis.title.y = element_text(size = 20),
    axis.title.x  = element_text(size = 20),
    legend.text = element_text(size = 20)
  )
p1


p2 <- input_data %>%
  as.data.frame() %>%
  mutate(claim_type = as.factor(claim_type)) %>%
  ggplot(aes(x = claim_type, fill = claim_type)) +
  geom_bar() +
  scale_fill_manual(
    values = c("royalblue", "#a71429"),
    labels = c("Claim type 0", "Claim type 1")
  ) +
  guides(fill = guide_legend(title = "Claim type")) +
  theme_bw() +
  labs(title = "", x = "Claim Type", y = "") +
  theme(
    axis.text = element_text(size = 20),
    axis.title.y = element_text(size = 20),
    axis.title.x  = element_text(size = 20),
    legend.text = element_text(size = 20)
  )
p2

Code for plotting feature-dependent development factors

Chain-Ladder case.

clmodel <- resurv_fit_cox$IndividualDataPP$training.data %>%
  mutate(DP_o = 16 -
           DP_rev_o + 1) %>%
  group_by(AP_o, DP_o) %>%
  summarize(I = sum(I), .groups = "drop") %>%
  group_by(AP_o) %>%
  arrange(DP_o) %>%
  mutate(I_cum = cumsum(I), I_cum_lag = lag(I_cum, default = 0)) %>%
  ungroup() %>%
  group_by(DP_o) %>%
  reframe(df_o = sum(I_cum * (
    AP_o <= max(resurv_fit_cox$IndividualDataPP$training.data$AP_o) - DP_o +
      1
  )) /
    sum(I_cum_lag * (
      AP_o <= max(resurv_fit_cox$IndividualDataPP$training.data$AP_o) - DP_o +
        1
    )),
  I = sum(I * (
    AP_o <= max(resurv_fit_cox$IndividualDataPP$training.data$AP_o) - DP_o
  ))) %>%
  mutate(DP_o_join = DP_o - 1) %>%
  as.data.frame()


clmodel %>%
  filter(DP_o > 1) %>%
  ggplot(aes(x = DP_o, y = df_o)) +
  geom_line(linewidth = 2.5,
            color = "#454555") +
  labs(title = "Chain ladder",
       x = "Development quarter",
       y = "Development factor") +
  ylim(1, 3. + .01) +
  theme_bw(base_size = rel(5)) +
  theme(plot.title = element_text(size = 20))

##

clmodel_months <- individual_data_m$training.data %>%
  mutate(DP_o = 48 -
           DP_rev_o + 1) %>%
  group_by(AP_o, DP_o) %>%
  summarize(I = sum(I), .groups = "drop") %>%
  group_by(AP_o) %>%
  arrange(DP_o) %>%
  mutate(I_cum = cumsum(I), I_cum_lag = lag(I_cum, default = 0)) %>%
  ungroup() %>%
  group_by(DP_o) %>%
  reframe(df_o = sum(I_cum * (
    AP_o <= max(individual_data_m$training.data$AP_o) - DP_o + 1
  )) /
    sum(I_cum_lag * (
      AP_o <= max(individual_data_m$training.data$AP_o) - DP_o + 1
    )),
  I = sum(I * (
    AP_o <= max(individual_data_m$training.data$AP_o) - DP_o
  ))) %>%
  mutate(DP_o_join = DP_o - 1) %>%
  as.data.frame()

ticks_at <- seq(1, 48, 4)
labels_as <- as.character(ticks_at)

clmodel_months %>%
  filter(DP_o > 1) %>%
  ggplot(aes(x = DP_o,
             y = df_o)) +
  geom_line(linewidth = 2.5,
            color = "#454555") +
  labs(title = "Chain ladder",
       x = "Development month",
       y = "Development factor") +
  ylim(1, 2.5 + .01) +
  scale_x_continuous(breaks = ticks_at, labels = labels_as) +
  theme_bw(base_size = rel(5)) +
  theme(plot.title = element_text(size = 20))


##

clmodel_years <- individual_data_y$training.data %>%
  mutate(DP_o = 4 -
           DP_rev_o + 1) %>%
  group_by(AP_o, DP_o) %>%
  summarize(I = sum(I), .groups = "drop") %>%
  group_by(AP_o) %>%
  arrange(DP_o) %>%
  mutate(I_cum = cumsum(I), I_cum_lag = lag(I_cum, default = 0)) %>%
  ungroup() %>%
  group_by(DP_o) %>%
  reframe(df_o = sum(I_cum * (
    AP_o <= max(individual_data_m$training.data$AP_o) - DP_o + 1
  )) /
    sum(I_cum_lag * (
      AP_o <= max(individual_data_m$training.data$AP_o) - DP_o + 1
    )),
  I = sum(I * (
    AP_o <= max(individual_data_m$training.data$AP_o) - DP_o
  ))) %>%
  mutate(DP_o_join = DP_o - 1) %>%
  as.data.frame()

ticks_at <- seq(1, 4, 1)
labels_as <- as.character(ticks_at)

clmodel_years %>%
  filter(DP_o > 1) %>%
  ggplot(aes(x = DP_o,
             y = df_o)) +
  geom_line(linewidth = 2.5,
            color = "#454555") +
  labs(title = "Chain ladder",
       x = "Development year",
       y = "Development factor") +
  ylim(1, 2.5 + .01) +
  scale_x_continuous(breaks = ticks_at, labels = labels_as) +
  theme_bw(base_size = rel(5)) +
  theme(plot.title = element_text(size = 20))

COX, quarterly output.

ct <- 0
ap <- 12

resurv_fit_predict_q$long_triangle_format_out$output_granularity %>%
  filter(AP_o == ap & claim_type == ct) %>%
  filter(row_number() == 1) %>%
  select(group_o)

plot(
  resurv_fit_predict_q,
  granularity = "output",
  title_par = "COX: Accident Quarter 12 Claim Type 0",
  group_code = 23
)


ct <- 0
ap <- 36

resurv_fit_predict_m$long_triangle_format_out$output_granularity %>%
  filter(AP_o == ap & claim_type == ct) %>%
  filter(row_number() == 1) %>%
  select(group_o)

plot(
  resurv_fit_predict_m,
  granularity = "output",
  color_par = "#a71429",
  title_par = "COX: Accident Month 36 Claim Type 0",
  group_code = 71
)



ct <- 1
ap <- 7

resurv_fit_predict_m$long_triangle_format_out$output_granularity %>%
  filter(AP_o == ap & claim_type == ct) %>%
  filter(row_number() == 1) %>%
  select(group_o)

plot(
  resurv_fit_predict_m,
  granularity = "output",
  color_par = "#a71429",
  title_par = "COX: Accident Month 7 Claim Type 1",
  group_code = 14
)


ct <- 0
ap <- 2

resurv_fit_predict_y$long_triangle_format_out$output_granularity %>%
  filter(AP_o == ap & claim_type == ct) %>%
  filter(row_number() == 1) %>%
  select(group_o)

plot(
  resurv_fit_predict_y,
  granularity = "output",
  color_par = "#FF6A7A",
  title_par = "COX: Accident Year 2 Claim Type 0",
  group_code = 3
)


ct <- 1
ap <- 3

resurv_fit_predict_y$long_triangle_format_out$output_granularity %>%
  filter(AP_o == ap & claim_type == ct) %>%
  filter(row_number() == 1) %>%
  select(group_o)


plot(
  resurv_fit_predict_y,
  granularity = "output",
  color_par = "#FF6A7A",
  title_par = "COX: Accident Year 3 Claim Type 1",
  group_code = 6
)

Code for computing ARE TOT and ARE CAL

The code below can be used to compute the performance metrics in Table 6 of the manuscript.

conversion_factor <- individual_data$conversion_factor

max_dp_i <- 1440

true_output <- resurv_fit_cox$IndividualDataPP$full.data %>%
  mutate(
    DP_rev_o = floor(max_dp_i * conversion_factor) -
      ceiling(
              DP_i * conversion_factor +
                ((AP_i - 1) %% (
                  1 / conversion_factor
                )) * conversion_factor) + 1,
    AP_o = ceiling(AP_i * conversion_factor),
    TR_o = AP_o - 1
  ) %>%
  filter(DP_rev_o <= TR_o) %>%
  group_by(claim_type, AP_o, DP_rev_o) %>%
  mutate(claim_type = as.character(claim_type)) %>%
  summarize(I = sum(I), .groups = "drop") %>%
  filter(DP_rev_o > 0)

out_list <- resurv_fit_predict_q$long_triangle_format_out
out <- out_list$output_granularity

#Total output
score_total <- out[, c("claim_type", "AP_o", "DP_o", "expected_counts")] %>%
  mutate(DP_rev_o = 16 - DP_o + 1) %>%
  inner_join(true_output, by = c("claim_type", "AP_o", "DP_rev_o")) %>%
  mutate(ave = I - expected_counts, abs_ave = abs(ave)) %>%
  # from here it is reformulated for the are tot
  ungroup() %>%
  group_by(AP_o, DP_rev_o) %>%
  reframe(abs_ave = abs(sum(ave)), I = sum(I))

are_tot <- sum(score_total$abs_ave) / sum(score_total$I)


dfs_output <- out %>%
  mutate(DP_rev_o = 16 - DP_o + 1) %>%
  select(AP_o, claim_type, DP_rev_o, f_o) %>%
  mutate(DP_rev_o = DP_rev_o) %>%
  distinct()

#Cashflow on output scale.Etc quarterly cashflow development
score_diagonal <- resurv_fit_cox$IndividualDataPP$full.data  %>%
  mutate(
    DP_rev_o = floor(max_dp_i * conversion_factor) -
      ceiling(
              DP_i * conversion_factor +
                ((AP_i - 1) %% (
                  1 / conversion_factor
                )) * conversion_factor) + 1,
    AP_o = ceiling(AP_i * conversion_factor)
  ) %>%
  group_by(claim_type, AP_o, DP_rev_o) %>%
  mutate(claim_type = as.character(claim_type)) %>%
  summarize(I = sum(I), .groups = "drop") %>%
  group_by(claim_type, AP_o) %>%
  arrange(desc(DP_rev_o)) %>%
  mutate(I_cum = cumsum(I)) %>%
  mutate(I_cum_lag = lag(I_cum, default = 0)) %>%
  left_join(dfs_output, by = c("AP_o", "claim_type", "DP_rev_o")) %>%
  mutate(I_cum_hat =  I_cum_lag * f_o,
         RP_o = max(DP_rev_o) - DP_rev_o + AP_o) %>%
  inner_join(true_output[, c("AP_o", "DP_rev_o")] %>%  distinct()
             , by = c("AP_o", "DP_rev_o")) %>%
  group_by(AP_o, DP_rev_o) %>%
  reframe(abs_ave2_diag = abs(sum(I_cum_hat) - sum(I_cum)), I = sum(I))

are_cal_q <- sum(score_diagonal$abs_ave2_diag) / sum(score_diagonal$I)


# scoring XGB ----

out_xgb <- resurv_predict_xgb$long_triangle_format_out$output_granularity

score_total_xgb <- out_xgb[, c("claim_type",
                               "AP_o",
                               "DP_o",
                               "expected_counts")] %>%
  mutate(DP_rev_o = 16 - DP_o + 1) %>%
  inner_join(true_output, by = c("claim_type", "AP_o", "DP_rev_o")) %>%
  mutate(ave = I - expected_counts, abs_ave = abs(ave)) %>%
  # from here it is reformulated for the are tot
  ungroup() %>%
  group_by(AP_o, DP_rev_o) %>%
  reframe(abs_ave = abs(sum(ave)), I = sum(I))

are_tot_xgb <- sum(score_total_xgb$abs_ave) / sum(score_total$I)


dfs_output_xgb <- out_xgb %>%
  mutate(DP_rev_o = 16 - DP_o + 1) %>%
  select(AP_o, claim_type, DP_rev_o, f_o) %>%
  mutate(DP_rev_o = DP_rev_o) %>%
  distinct()

#Cashflow on output scale.Etc quarterly cashflow development
score_diagonal_xgb <- resurv_fit_cox$IndividualDataPP$full.data  %>%
  mutate(
    DP_rev_o = floor(max_dp_i * conversion_factor) -
      ceiling(
              DP_i * conversion_factor +
                ((AP_i - 1) %% (
                  1 / conversion_factor
                )) * conversion_factor) + 1,
    AP_o = ceiling(AP_i * conversion_factor)
  ) %>%
  group_by(claim_type, AP_o, DP_rev_o) %>%
  mutate(claim_type = as.character(claim_type)) %>%
  summarize(I = sum(I), .groups = "drop") %>%
  group_by(claim_type, AP_o) %>%
  arrange(desc(DP_rev_o)) %>%
  mutate(I_cum = cumsum(I)) %>%
  mutate(I_cum_lag = lag(I_cum, default = 0)) %>%
  left_join(dfs_output_xgb, by = c("AP_o", "claim_type", "DP_rev_o")) %>%
  mutate(I_cum_hat =  I_cum_lag * f_o,
         RP_o = max(DP_rev_o) - DP_rev_o + AP_o) %>%
  inner_join(true_output[, c("AP_o", "DP_rev_o")] %>%  distinct()
             , by = c("AP_o", "DP_rev_o")) %>%
  group_by(AP_o, DP_rev_o) %>%
  reframe(abs_ave2_diag = abs(sum(I_cum_hat) - sum(I_cum)), I = sum(I))

are_cal_q_xgb <- sum(score_diagonal_xgb$abs_ave2_diag) / sum(score_diagonal$I)

# scoring NN ----

out_nn <- resurv_predict_nn$long_triangle_format_out$output_granularity

score_total_nn <- out_nn[, c("claim_type",
                             "AP_o",
                             "DP_o",
                             "expected_counts")] %>%
  mutate(DP_rev_o = 16 - DP_o + 1) %>%
  inner_join(true_output, by = c("claim_type", "AP_o", "DP_rev_o")) %>%
  mutate(ave = I - expected_counts, abs_ave = abs(ave)) %>%
  # from here it is reformulated for the are tot
  ungroup() %>%
  group_by(AP_o, DP_rev_o) %>%
  reframe(abs_ave = abs(sum(ave)), I = sum(I))

are_tot_nn <- sum(score_total_nn$abs_ave) / sum(score_total$I)


dfs_output_nn <- out_nn %>%
  mutate(DP_rev_o = 16 - DP_o + 1) %>%
  select(AP_o, claim_type, DP_rev_o, f_o) %>%
  mutate(DP_rev_o = DP_rev_o) %>%
  distinct()

#Cashflow on output scale.Etc quarterly cashflow development
score_diagonal_nn <- resurv_fit_cox$IndividualDataPP$full.data  %>%
  mutate(
    DP_rev_o = floor(max_dp_i * conversion_factor) -
      ceiling(
              DP_i * conversion_factor +
                ((AP_i - 1) %% (
                  1 / conversion_factor
                )) * conversion_factor) + 1,
    AP_o = ceiling(AP_i * conversion_factor)
  ) %>%
  group_by(claim_type, AP_o, DP_rev_o) %>%
  mutate(claim_type = as.character(claim_type)) %>%
  summarize(I = sum(I), .groups = "drop") %>%
  group_by(claim_type, AP_o) %>%
  arrange(desc(DP_rev_o)) %>%
  mutate(I_cum = cumsum(I)) %>%
  mutate(I_cum_lag = lag(I_cum, default = 0)) %>%
  left_join(dfs_output_nn, by = c("AP_o", "claim_type", "DP_rev_o")) %>%
  mutate(I_cum_hat =  I_cum_lag * f_o,
         RP_o = max(DP_rev_o) - DP_rev_o + AP_o) %>%
  inner_join(true_output[, c("AP_o", "DP_rev_o")] %>%  distinct()
             , by = c("AP_o", "DP_rev_o")) %>%
  group_by(AP_o, DP_rev_o) %>%
  reframe(abs_ave2_diag = abs(sum(I_cum_hat) - sum(I_cum)), I = sum(I))

are_cal_q_nn <- sum(score_diagonal_nn$abs_ave2_diag) / sum(score_diagonal$I)

# Scoring Chain-Ladder ----

true_output_cl <- individual_data$full.data %>%
  mutate(
    DP_rev_o = floor(max_dp_i * conversion_factor) -
      ceiling(
              DP_i * conversion_factor +
                ((AP_i - 1) %% (
                  1 / conversion_factor
                )) * conversion_factor) + 1,
    AP_o = ceiling(AP_i * conversion_factor)
  ) %>%
  filter(DP_rev_o <= TR_o) %>%
  mutate(DP_o = max(individual_data$training.data$DP_rev_o) - DP_rev_o + 1) %>%
  group_by(AP_o, DP_o, DP_rev_o) %>%
  summarize(I = sum(I), .groups = "drop") %>%
  filter(DP_rev_o > 0)
latest_observed <- individual_data$training.data %>%
  filter(DP_rev_o >= TR_o) %>%
  mutate(DP_o = max(individual_data$training.data$DP_rev_o) - DP_rev_o + 1) %>%
  group_by(AP_o) %>%
  mutate(DP_max = max(DP_o)) %>%
  group_by(AP_o, DP_max) %>%
  summarize(I = sum(I), .groups = "drop")

clmodel <- individual_data$training.data %>%
  mutate(DP_o = max(individual_data$training.data$DP_rev_o) - DP_rev_o + 1) %>%
  group_by(AP_o, DP_o) %>%
  summarize(I = sum(I), .groups = "drop") %>%
  group_by(AP_o) %>%
  arrange(DP_o) %>%
  mutate(I_cum = cumsum(I), I_cum_lag = lag(I_cum, default = 0)) %>%
  ungroup() %>%
  group_by(DP_o) %>%
  reframe(df = sum(I_cum * (
    AP_o <= max(individual_data$training.data$AP_o) - DP_o + 1
  )) /
    sum(I_cum_lag * (
      AP_o <= max(individual_data$training.data$AP_o) - DP_o + 1
    )), I = sum(I * (
    AP_o <= max(individual_data$training.data$AP_o) - DP_o
  ))) %>%
  mutate(DP_o_join = DP_o) %>%
  mutate(DP_rev_o = max(DP_o) - DP_o + 1)

predictions <- expand.grid(AP_o = latest_observed$AP_o,
                           DP_o = clmodel$DP_o_join) %>%
  left_join(clmodel[, c("DP_o_join", "df")], by = c("DP_o" = "DP_o_join")) %>%
  left_join(latest_observed, by = "AP_o") %>%
  rowwise() %>%
  filter(DP_o > DP_max) %>%
  ungroup() %>%
  group_by(AP_o) %>%
  arrange(DP_o) %>%
  mutate(df_cum = cumprod(df)) %>%
  mutate(I_expected = I * df_cum - I * lag(df_cum, default = 1)) %>%
  select(DP_o, AP_o, I_expected)

conversion_factor <- individual_data$conversion_factor
max_dp_i <- 1440
score_total <- predictions  %>%
  inner_join(true_output_cl, by = c("AP_o", "DP_o")) %>%
  mutate(ave = I - I_expected, abs_ave = abs(ave)) %>%
  # from here it is reformulated for the are tot
  ungroup() %>%
  group_by(AP_o, DP_rev_o) %>%
  reframe(abs_ave = abs(sum(ave)), I = sum(I))

are_tot <- sum(score_total$abs_ave) / sum(score_total$I)

score_diagonal <- individual_data$full.data  %>%
  mutate(
    DP_rev_o = floor(max_dp_i * conversion_factor) -
      ceiling(
              DP_i * conversion_factor +
                ((AP_i - 1) %% (
                  1 / conversion_factor
                )) * conversion_factor) + 1,
    AP_o = ceiling(AP_i * conversion_factor)
  ) %>%
  group_by(claim_type, AP_o, DP_rev_o) %>%
  mutate(claim_type = as.character(claim_type)) %>%
  summarize(I = sum(I), .groups = "drop") %>%
  group_by(claim_type, AP_o) %>%
  arrange(desc(DP_rev_o)) %>%
  mutate(I_cum = cumsum(I)) %>%
  mutate(I_cum_lag = lag(I_cum, default = 0)) %>%
  mutate(DP_o = max(individual_data$training.data$DP_rev_o) - DP_rev_o + 1) %>%
  left_join(CL[, c("DP_o", "df")], by = c("DP_o")) %>%
  mutate(I_cum_hat =  I_cum_lag * df,
         RP_o = max(DP_rev_o) - DP_rev_o + AP_o) %>%
  inner_join(true_output_cl[, c("AP_o", "DP_rev_o")] %>%  distinct()
             , by = c("AP_o", "DP_rev_o")) %>%
  group_by(AP_o, DP_rev_o) %>%
  reframe(abs_ave2_diag = abs(sum(I_cum_hat) - sum(I_cum)), I = sum(I))

are_cal_q <- sum(score_diagonal$abs_ave2_diag) / sum(score_diagonal$I)

Claim Counts Prediction Using Individual Data

Gabriele Pittarello

2024-11-14