Original subtitles

So in this lecture, we are going to discuss the types of tasks that you can do with Time series models.

We'll start with the most intuitive task, which for most people is the one step forecast that is given.

Why one and why two all the way up to what we would like to predict, YFC plus one.

And so that's what we train our model to do.

As a side note, be aware that in this class we might use X of T or Y of T to represent a time series.

So in the context of a long time series, there won't be any difference between using X or Y in a supervised

learning.

This is a different story.

So just keep this in the back of your mind.

Now, the one step forecast is not the end of the story in practice, we often want to predict multiple

steps into the future.

We call this the forecast horizon.

This is the length of time into the future that we would like to predict.

For example, we might want to predict the sales of each day for the next week.

One very realistic example is the weather.

Imagine checking the Weather Channel and it only shows the weather one day ahead.

This wouldn't be very useful compared to other weather channels.

Now, this is not to say that one step forecasts are not useful.

It all depends on context.

For example, if you run a brick and mortar shop, you might want to forecast the sales of our products

next month from monthly data.

In this case, that is a one step forecast.

This will help you purchase any inventory you might need to fulfill demand in the next month.

And again, this all depends on context.

When in doubt, ask your manager or your clients or your stakeholders.

So in this course, we're going to learn two ways in which to produce multi-step forecast method number

one is called the incremental forecast, which can be done with any one step predictor.

Method number two is called the multi output forecast, which is limited to only certain models.

For example, we'll study Arima, which can do one step forecasts but cannot do a multi output forecast.

So let's discuss each of these models more in depth.

For now, suppose that our model is a black box.

It can only make a one step forecast.

Let's suppose that in order to make this prediction, we use past data points.

So given we of T minus P plus one up to Y of T, we can predict Y of T plus one.

Let's call this prediction Y hat of T plus one.

But suppose that our forecast horizon is three times steps so H is equal to three.

In this case, what we can do is plug in our prediction as an input into our model.

So for the next input will pass in Y of T minus plus two up to Y of T followed by Y had of T plus one.

This will give us Y hat of T plus two.

Now that we have our prediction for Y had of T plus two, we can take this, plug it into our input

again and yet we have T plus three.

The important thing to recognize is that we are not allowed to use the true values for YFC plus one

and YFC plus two because the current time is only T, C plus one and C plus two are in the future.

So just to give you a concrete example, suppose that we have we have one, we have two and we have

three say that is equal to three.

So our model uses three pass time points to predicts the next point.

We'd like to predict why.

Six.

So our forecast horizon is three.

Before we can do that, we must find we had four.

So we use our model plugging in.

Why one, why to and why three.

And that gives us we had four.

Now in order to estimate why five we plug in.

Why two.

Why three.

And why had four.

This gives us why had five.

Finally we can plug in.

Why three.

We had four and we had five.

And this will give us we had six.

Now you might be wondering why can't you just plug in the true values.

Why four and why five in order to estimate why six.

The answer is if you imagine these are days to day is only day three in order to know the true value

of why four and why five, we would have to wait until day five.

And of course, in some businesses this might be unacceptable.

As I always tell students who ask this question, imagine your boss asks you to make a forecast three

days ahead.

Your answer cannot be OK, but we just have to wait a two days to get an answer.

As always, context is important.

So ask your clients or your manager what your forecast horizon really is.

Now, let's talk about method number two, method number two is the multi output forecast.

Again, imagine that we have a black box model for some models.

We will study.

We'll see that they are capable of giving us multiple outputs at once.

Therefore, if we have a forecast to rise Horizon H, then we can simply create a model with H outputs

and obtain a forecast for each time.

Step up to points in the future.

As you can see, this is much simpler than building a forecast incrementally.

The final task I want to discuss in this lecture is that of classification, this isn't normally discussed

in a traditional time series analysis, but this being a more modern course, it's worth knowing.

So in the previous examples, our job was always to predict a number in machine learning.

We call this regression, but what if we'd like to predict a category instead?

Again, let's pretend that we are neural link.

We want to read a user's brain signals and yes, whether they are hungry or tired, hungry and tired,

our categories, not numbers.

Another example is taking motion readings from a smartphone and trying to guess what the user is doing

walking, sleeping or sitting down.

Again, these are categories, not numbers.

OK, so these are some examples of the types of tasks that we can do with Time series analysis.

Thanks for listening and I'll see you in the next lecture.