new stock experiment and bug fixes

Author: dkesada

NAMESPACE

@@ -1,6 +1,7 @@
 # Generated by roxygen2: do not edit by hand
 
 export(main_prep_and_run_motor)
+export(main_prep_and_run_stock)
 export(main_prep_and_run_synth)
 export(mtDBN)
 export(treeSc)

R/mtDBN.R

@@ -326,15 +326,16 @@ mtDBN <- R6::R6Class("mtDBN",
 
       # Operate the M5 smoothing up to the root node
       node <- private$tree_sc$classify_inst(instance)
-
-      #while(is.environment(node$p_node)){
+      
+      #while(is.environment(node$p_node)){ # In case an smoothing model is built per split point, which is very costly
+      if(is.environment(node$p_node)){
         parent_model <- private$models[[as.character(node$p_node$name)]]
         preds_p <- private$exact_prediction_step(parent_model, vars_pred,
                                                  private$as_named_vector(instance[1, .SD, .SDcols = c(vars_ev, prov_ev)]))
         preds$mu_p <- Map(function(x,y){(node$n * x + k * y) / (node$n + k)}, preds$mu_p, preds_p$mu_p)
         #preds$mu_p <- Map(function(x,y){(x + y) / (2)}, preds$mu_p, preds_p$mu_p)
         node <- node$p_node
-      #}
+      }
 
       return(preds)
     }

R/stock_exp.R

@@ -0,0 +1,21 @@
+#' @export
+main_prep_and_run_stock <- function(){
+  dt <- data.table::fread("./dataset/TWII_1y.csv") # 0.5 secs between rows
+  id_var <- "Date" # We will crossvalidate based on the months
+  
+  dt[, Date := unlist(strsplit(as.character(dt$Date), "-[0-9]*$"))] # We leave only the year and month of the dates for making CV groups
+  dt[, Volume := as.numeric(Volume)]
+  setnames(dt, "Adj Close", "Adj_Close")
+  obj_vars <- "Open_t_0"
+  obj_var_univ <- "Open"
+  obj_var_multiv <- c("Open", "Volume")
+  prov_ev <- NULL
+  # max_min_vals <- max_min_norm(dt, names(dt)[-1])
+  dt[, Adj_Close := NULL] # Exactly the same as Close
+    
+  full_exp_run(dt = dt, id_var = id_var, obj_vars = obj_vars,
+               obj_var_univ = obj_var_univ, obj_var_multiv = obj_var_multiv,
+               prov_ev = prov_ev, res_file = "full_run_stock_results.txt",
+               mae_file = "full_run_stock_mae.csv", mape_file = "full_run_stock_mape.csv", pred_len = 1,
+               fold_len = 2, seed = 42, size = 3, method = "psoho", n_it = 100, min_ind = 50)
+}

R/utils.R

@@ -223,6 +223,16 @@ mape <- function(orig, pred){
   return((100/length(orig)) * sum(abs((orig - pred) / orig)))
 }
 
+max_min_norm <- function(dt, cols = names(dt)){
+  min_vals <- dt[, sapply(.SD, min), .SDcols = cols]
+  max_vals <- dt[, sapply(.SD, max), .SDcols = cols]
+  
+  for(v in cols)
+    dt[, (v) := (get(v) - min_vals[v]) / (max_vals[v] - min_vals[v])]
+  
+  return(list(max_vals = max_vals, min_vals = min_vals))
+}
+
 # Experiment pipeline functions
 
 forecast_cycle_intervals_single <- function(f_dt_test, model_fit, id_var, size, obj_vars, prov_ev, pred_len){