Streamline Your Document Processes with airSlate SignNow's B2B Sales Forecasting for R&D

How to create outlook signature

hey guys welcome to arc tip 40 and we are on the 40th r tip and i wanted to focus on a package that i've actually built uh and it's an ecosystem of forecasting libraries and i'm going to teach you how to use it it's called model time and we're going to introduce you to it it's going to be a gentle introduction and we're going to learn it over the course of about 10 minutes and we're actually going to create several forecasts three different forecasts two of them will be really good and it'll show you the value of the model time ecosystem so to get started here what you need to do is if you want the code that we're going over today just sign up for our business science r tips newsletter there's a link in the video notes and what that'll do is give you access to our github repo we'll do a get pool once you pull your latest files and folders in you will have 40 different r tips here and we'll be focusing on the 40th one model time forecasting just click this and open up this dot r file that's what we're going to be working out of okay and again what we're going to be doing is making this forecast here so let's get started first thing i'm going to do is load in the libraries so i'm going to do control and enter and what this does is it loads in model time and the tidy models which these two work together for modeling time series in our using the latest tools and techniques uh in machine learning and then you'll load tidy verse and time tk time tk is another package that i've built it's for time series data manipulations and then the tidy versus one of our core libraries that we always load in comes with it just pre packages dplyr ggplot2 and several other libraries that are very important we're also going to be working with date features so the lubridate package is one of the fundamental or foundational time series packages we'll be using month and weekday functions from that library all right next what we're going to do is we're going to load in the bike sharing daily it's a time series data set that comes with time tk and it's 731 observations it's a daily time series so you can see there's time stamps oh 101.02 and then what we're going to be focusing on is this count and what that is is the bike shares the number of people that are sharing bikes and using bikes in dc during this time period okay so we're just going to select the dte and count columns and that's going to be stored as a data frame called data so we'll just take a quick look at data it should just be two columns and here it is a table 731 by 2 with just the date and the count we can get a quick plot of it using this function from plot time series and the x-axis is going to be dte day the count is going to be the y-axis and we can get a clear picture of how this time series works we see that over the year there's some sort of seasonality but there's also an increasing trend so we want to use some time series forecasting tools to be able to forecast what's going to happen over the next three months so what we're going to do is we're first going to split into a training testing set and this is what time series splits does this what this will do is create a splits which object which breaks it up temporally meaning using the last three months of the time series and we can visualize that with these next few lines of code all we're doing is extracting out the time series cross validation plan and then visualizing it with dte day and count and we can see that there's a training and testing set the blue region is going to be our training period the red region is going to be our testing period and this is what we're going to assess the model performance with okay once we have that we're ready to start forecasting so if you're following along in the outline we're right now here we're going to forecast with profit auto arima is going to be our first one so that should show up here and this is how it works uh basically model time comes with an arima rig reg function and this allows us to specify a model we're going to set the engine to auto arima and what this does is it connects up to the forecast library for using the auto arima function and then we're going to fit that and it's going to be count as a function of the dte day so once we fit that model it creates a model for us now this may seem like a lot of work but this is actually uh very concise and it's able to be and it's consistent which means we're able to kind of do the same thing when we move on to our next modeling approach so again we're going to work on a new model now profit but the approach is is basically the same thing we first create the um the profit regression specification uh here i'm passing an argument telling it to look at the seasonal the seasonality yearly i'm going to set the engine to profit and this that connects it up to the profit library for this regression model and then i'm going to fit that model so when i run all of this it creates a profit model for me and the nice thing is is that it would take you a lot of time if you're working in the forecast library and the profit library because they each take their data in the same different ways so this makes everything super consistent and concise okay and then the third thing so now we'll even try some machine learning so we're going to do what's called a glm and it's from the glm net library so we're going to set up a specification for a linear regression and this then we're going to supply a parameter called penalty which is penalized regression and then we're going to set the engine to glm-net which is uh for the for for the elastic net penalty penalized regression we're going to fit it and here i'm going to expand on our kind of formula process so in the fit it takes a formula and i'm going to specify count as a function of the weekday with it labeled equal to true i'm going to add in a month feature which is going to be just the month whether or not it's january february and so on and then i'm also going to add in a trend feature so i'm going to convert that date to a number using the as numeric function so when i run this we get another model and it's basically the same process so now we have three models already and it's all the same consistent process and now here's where model time really takes over so now we can start to first organize so we're going to create a model time table and we just use the model time table function so if i run this control enter we now have a model table which stores each of our three different models here and it also gives them a description and an id feature so just basically for organizing all three of our models we're then going to calibrate them so what calibration does is it assesses the residuals on the testing set so you see here we have the testing set so you're going to see two new columns that get added to our model timetable we have a table here of 90 by 4 and this is calibration data so it's basically looking at that test data set which is this red region here which is if we see here it's 90 observations so we're actually storing the residuals in here and it's calculating the actual versus the prediction and just storing those for each of our models and this is what allows us to then move on and collect the accuracy so we can very quickly get the model time accuracy and these are different measures of accuracy here so mean absolute error is a very popular one we can see arima is is performing the worst uh profit looks like it's performing the best and glm that's performing right around where profit's at we can see that from an r squared which measures variability that looks like profit is doing one of the best so profit is it looks like it's going to be a pretty strong model in this instance okay next what we're going to do is visualize and this is always a good idea so we can use another function called model time forecast and we're going to provide the testing splits and then we're also going to use the actual data to overlay or to underlay and i'll show you what this does when we take the output of model time forecasts and run it through plot model time forecasts we get an interactive plot and we can more visually be able to see what's going on with the specific models so we can see here that arima doesn't look like it's doing very well it's getting the overall trend but it's not picking up the seasonality the yearly seasonality conversely profit and geolemna both seem to be getting the yearly seasonality correct so if i take this one off i can actually zoom in and take a closer look at these two models and see what they look like okay all right the next thing that we can do is now that we understand a little bit more about our models we can now forecast the future so we can take that same calibration data and the only major difference here is we're refitting so when we refit that actually retrains all three of these models on the full data set so before they were just trained on the training portion which ends here now we're going to train it on the full data set and then we're going to forecast it for the next three months so we do that and then we can plot our model time forecast and here is the final forecast again we're going to take off the arima model because that one didn't do so hot and this is what the profit and glm net both look like okay and in under 10 minutes we have just successfully given you your first introduction to model time there's a lot more to it so if you're interested in learning more these models specifically there's a lot of room for improvement uh there's things that we can do so there's feature engineering with the recipes library this is a big area where we can begin to add in things like lags things like additional trend features for your series and more in the recipes library i teach that in my time series 203 course and i'll talk more about that here in a second there's also machine learning algorithms so we've got more algorithms that we can use than just what we've showcased here we've showcased auto arima profit and in the glm but there's things like xg boost there's things like um random forests there's even more specialized models like luan ts so this model time ecosystem is a growing ecosystem and you can imagine that we're adding more and more sophistication to it two libraries i want to point to model time h2o and gluon ts so i teach all of these in my high performance time series course if you're interested in that i have a link here and it's also in the chat definitely check those out these will help you do forecasting at scale with many many time series and it's basically all about developing a high performance forecasting system so check it out if you're interested in the links in the video notes alright see you next time