1
00:00:00,480 --> 00:00:06,600
Sometimes when we created time, lot of our variable, we see a plot like this.

2
00:00:07,590 --> 00:00:11,940
That is the overall trend in our variable is non-linear.

3
00:00:13,830 --> 00:00:22,290
For example, if you have product sales data and your product suddenly goes viral, it will show exponential

4
00:00:22,290 --> 00:00:23,980
growth as compared to linear growth.

5
00:00:26,130 --> 00:00:33,270
But the problem is that a lot of our forecasting models do not work well when there is a nonlinear trend.

6
00:00:35,880 --> 00:00:41,430
Although we haven't started discussing models yet, I'm covering this topic now because power transformation

7
00:00:41,670 --> 00:00:44,520
is also a step we take during data processing.

8
00:00:46,560 --> 00:00:49,640
Some models may not require it, but some models do.

9
00:00:50,730 --> 00:00:52,110
And the concept is very simple.

10
00:00:53,280 --> 00:00:55,440
Let me show you what our transformation is.

11
00:01:00,630 --> 00:01:06,420
So suppose you had these values of the variable and when we plotted, we see a line just like this.

12
00:01:07,620 --> 00:01:09,270
Clearly, this is not a linear trend.

13
00:01:10,920 --> 00:01:13,860
So you use your knowledge of mathematical functions.

14
00:01:14,340 --> 00:01:16,560
And guess what this function looks like?

15
00:01:18,180 --> 00:01:23,310
For example, this particular graph looks like an exponential function.

16
00:01:25,950 --> 00:01:31,810
So the concept here, a simple sense, these values of the variable show exponential trend.

17
00:01:33,060 --> 00:01:38,880
If you take a log of these values, those values will then show linear trend.

18
00:01:40,860 --> 00:01:44,040
And on the right is a plot of logged values.

19
00:01:45,630 --> 00:01:49,530
You can clearly see that this plot is more linear.

20
00:01:52,350 --> 00:01:56,550
So we can use these logged values in our model for forecasting.

21
00:01:58,470 --> 00:02:04,260
Similarly, data can have quadratic logarithmic or any polynomial distribution.

22
00:02:06,180 --> 00:02:12,430
If you sense such a relationship, convert the values correspondingly so that you get a more lenient

23
00:02:12,430 --> 00:02:12,720
thing.

24
00:02:15,540 --> 00:02:19,260
Again, this is not a mandatory step for forecasting models.

25
00:02:19,950 --> 00:02:23,170
Only some models perform better if the trend is linear.

26
00:02:24,780 --> 00:02:30,360
When we discuss models, I will highlight which are those models with power transformation can help

27
00:02:30,450 --> 00:02:31,470
improve the predictions.