library(CASdatasets)
library(tidyverse)
data("norauto")

sum(norauto$ClaimAmount)
sum(norauto$Expo)

sum(norauto$ClaimAmount)/sum(norauto$Expo)
norauto %>% group_by(GeoRegion) %>% summarise(media = sum(ClaimAmount)/sum(Expo))

norauto$ClaimAmount2 <- ifelse(norauto$ClaimAmount>=50000, 50000, norauto$ClaimAmount)
factor_elevacion = sum(norauto$ClaimAmount)/sum(norauto$ClaimAmount2)
norauto$ClaimAmount2 = norauto$ClaimAmount2 * factor_elevacion

norauto$prima_pura = norauto$ClaimAmount / norauto$Expo


norauto$log_exposure <- log(norauto$Expo)
norauto2 <- norauto %>% sample_frac(0.1)

predictoras <- c('Male','Young','DistLimit','GeoRegion')
respuesta <- "ClaimAmount2"
offset_var <- "log_exposure"

library(h2o)
h2o.init()
#h2o.shutdown()

norauto.hex <- as.h2o(norauto2)

auto.splits <- h2o.splitFrame(data = norauto.hex, ratios = .6)
train <- auto.splits[[1]]
valid <- auto.splits[[2]]

parametro_tweedie <- list( tweedie_variance_power = c(1.5,1.6,1.7,1.8,1.9,2,2.2,2.5))

grid <- h2o.grid(seed=10,
                 x = predictoras, y = respuesta, training_frame = train, validation_frame = valid,
                 family = 'tweedie', algorithm = "glm", grid_id = "auto_grid",
                 hyper_params = parametro_tweedie, 
                 search_criteria = list(strategy = "Cartesian"))

busqueda_ordenada <- h2o.getGrid("auto_grid", sort_by = "mse", decreasing = F)

vp = as.numeric(busqueda_ordenada@summary_table[1,1])
print(vp)

ntrees_opts = c(500)       
max_depth_opts = seq(5,7)
min_rows_opts = c(50,100)
learn_rate_opts = c(0.01)
sample_rate_opts = seq(0.5,0.75)
col_sample_rate_opts = seq(0.5,0.8)
col_sample_rate_per_tree_opts = seq(0.5,1)
nbins_cats_opts = seq(100,500) 

hyper_params = list( ntrees = ntrees_opts,
                     max_depth = max_depth_opts,
                     min_rows = min_rows_opts,
                     learn_rate = learn_rate_opts,
                     sample_rate = sample_rate_opts,
                     col_sample_rate = col_sample_rate_opts,
                     col_sample_rate_per_tree = col_sample_rate_per_tree_opts,
                     nbins_cats = nbins_cats_opts)

search_criteria = list(strategy = "RandomDiscrete",
                       max_runtime_secs = 600,
                       max_models = 100,
                       stopping_metric = "AUTO",
                       stopping_tolerance = 0.00001,
                       stopping_rounds = 5,
                       seed = 123456)


gbm_grid <- h2o.grid("gbm",
                     grid_id = "mygrid",
                     x = predictoras,
                     y = respuesta,
                     tweedie_power=vp,
                     offset_column=offset_var,
                     training_frame = train,
                     validation_frame = valid,
                     nfolds = 0,
                     distribution="tweedie",
                     stopping_rounds = 2,
                     stopping_tolerance = 1e-3,
                     stopping_metric = "MSE",
                     score_tree_interval = 100,
                     seed = 123456,
                     hyper_params = hyper_params,
                     search_criteria = search_criteria)

sorted_grid <- h2o.getGrid(grid_id = "mygrid", sort_by = "mse")

modelo_final <- h2o.getModel(sorted_grid@model_ids[[1]])

norauto.hex.total <- as.h2o(norauto)
estimaciones <- as.data.frame(predict(modelo_final,norauto.hex.total))
names(estimaciones) <- "prima_estimada"

norauto <- cbind.data.frame(norauto,estimaciones)

sum(norauto$ClaimAmount)/sum(norauto$prima_estimada)-1