marked the best number of observations when doing observation based modelling and implemented fullforecast.serial.obs

Author: paulbarbu

Diff for: 1hrs_ph3_report_generator.R

@@ -224,11 +224,12 @@ for(testobs in 1:(frequency(datasets[['1hrs ph3']]$series) - 1)) # 1:23
   }
 }
 
+#RMSE 336
 report.full(model = 'Arima(order=c(1, 0, 0), seasonal=c(1, 0, 0), method="CSS")',
             series = '1hrs ph3',
             transformation = 'identity()',
             traindays = trainobs,
-            testdays = best.testobs, # ?
+            testdays = best.testobs, # 1
             obs = TRUE)
 
 
@@ -252,10 +253,11 @@ for(testobs in 1:(frequency(datasets[['1hrs ph3']]$series) - 1)) # 1:11
   }
 }
 
+#RMSE 316
 report.full(model = 'Arima(order=c(1, 0, 0), seasonal=c(1, 0, 0), xreg=fourier(., K=2), method="CSS")',
             series = '1hrs ph3',
             transformation = 'identity()',
             traindays = trainobs,
-            testdays = best.fourier.testobs, # ?
-            xreg = 'fourier(., h=h, K=1)',
-            obs = TRUE)
+            testdays = best.fourier.testobs, # 1
+            xreg = 'fourier(., h=h, K=2)',
+            obs = TRUE)

Diff for: funcs.R

@@ -337,3 +337,73 @@ fullforecast.obs <- function(dataset, transformation, model, trainobs, testobs,
 
   return(fcasts)
 }
+
+
+fullforecast.serial.obs <- function(dataset, transformation, model, trainobs, testobs, xreg, max.iterations = NULL)
+{
+  startobs <- 0
+  
+  if(is.null(max.iterations))
+  {
+    endobs <- length(dataset) - (trainobs + testobs)
+  }else{
+    endobs <- max.iterations
+  }
+  
+  fcasts <- list()
+  fcasts$points <- c()
+  
+  for(currentobs in startobs:endobs)
+  {
+    print(paste("Current obs =", currentobs))
+    datachunk <- get_obs(dataset, currentobs, trainobs, testobs)
+    
+    eval(parse(text=paste('datachunk$train %>%',
+                          transformation))) -> datachunk$train.transformed
+    
+    h <- testobs
+    
+    # for naive, snaive and meanf models
+    # the model already returns the forecasts, no separate forecast step
+    # is needed
+    if(any(startsWith(model, c("naive()", "snaive()", "meanf()"))))
+    {
+      model <- sub("()", paste("(h=", h, ")", sep=""), model, fixed=TRUE)
+    } 
+    
+    eval(parse(text=paste('datachunk$train.transformed %>%', model))) -> fit
+    
+    fcast <- NULL
+    
+    # for naive, snaive and meanf models
+    # the model already returns the forecasts, no separate forecast step
+    # is needed
+    if(any(startsWith(model, c("naive(", "snaive(", "meanf("))))
+    {
+      chunk.fcast <- fit
+    } else
+    {
+      fourier.terms <- NULL
+      
+      if(!is.null(xreg))
+      {
+        eval(parse(text=paste('datachunk$test %>%', xreg))) -> fourier.terms
+      }
+      
+      fit %>% forecast(h=h, xreg=fourier.terms) -> chunk.fcast
+    }
+    
+    fcasts$points <- c(fcasts$points, chunk.fcast$mean)
+    
+    gc()
+  }
+  
+  gc()
+  
+  fcasts$points <- ts(fcasts$points, start=trainobs/frequency(dataset), frequency = frequency(dataset))
+  testpoints <- get_obs(dataset, trainobs, NULL, NULL)$train
+  
+  fcasts$accuracy <- accuracy(fcasts$points, testpoints)
+  
+  return(fcasts)
+}

Diff for: mem_profile.R

@@ -1,4 +1,4 @@
-library(lineprof)
+library(profvis)
 
 source("funcs.R")
 prof.2hrsph3 <- lineprof(
@@ -9,7 +9,7 @@ prof.2hrsph3 <- lineprof(
                       testdays = 1,
                       xreg = NULL,
                       max.iterations = 2))
-shine(prof.2hrsph3)
+print(prof.2hrsph3)
 
 prof.ph3 <- lineprof(
   fullforecast.serial(model = 'Arima(order=c(1, 0, 0), seasonal=c(1, 0, 0), method="ML")',
@@ -19,4 +19,28 @@ prof.ph3 <- lineprof(
                       testdays = 1,
                       xreg = NULL,
                       max.iterations = 2))
-shine(prof.ph3)
+print(prof.ph3)
+
+prof.ph3.serial.obs <- profvis(
+    fullforecast.serial.obs(model = 'Arima(order=c(1, 0, 0), seasonal=c(1, 0, 0), method="ML")',
+                        dataset = datasets[['ph3']]$series,
+                        transformation = 'identity()',
+                        trainobs = 12, #frequency(datasets[['ph3']]$series),
+                        testobs = 1,
+                        xreg = NULL,
+                        max.iterations = 0)
+)
+
+print(prof.ph3.serial.obs)
+
+
+prof.2hrsph3.serial.obs <- profvis(
+  fullforecast.serial.obs(model = 'Arima(order=c(1, 0, 0), seasonal=c(1, 0, 0), method="ML")',
+                        dataset = datasets[['2hrs ph3']]$series,
+                        transformation = 'identity()',
+                        trainobs = 12, #frequency(datasets[['ph3']]$series),
+                        testobs = 1,
+                        xreg = NULL,
+                        max.iterations = 0)
+)
+print(prof.2hrsph3.serial.obs)