1 00:00:00,430 --> 00:00:04,180 We'll go back to the simple DNN that we saw way back in week two for 2 00:00:04,180 --> 00:00:06,618 training on and we'll see what happens. 3 00:00:06,618 --> 00:00:11,559 We get a chart like this, which at least to the eyeball looks really good, 4 00:00:11,559 --> 00:00:15,157 but it has a very large MAE so something must be wrong. 5 00:00:15,157 --> 00:00:20,071 Indeed, if we zoom into the results we can see in a little bit more detail about 6 00:00:20,071 --> 00:00:23,109 how the forecast behaves in the original data. 7 00:00:23,109 --> 00:00:26,690 Our clue to the problem could be our window size. 8 00:00:26,690 --> 00:00:28,700 Remember earlier we said it's a 20 so 9 00:00:28,700 --> 00:00:32,670 our training window sizes are 20 time slices worth of data. 10 00:00:32,670 --> 00:00:35,820 And given that each time slice is a month in real time 11 00:00:35,820 --> 00:00:38,270 our window is a little under two years. 12 00:00:39,360 --> 00:00:40,950 But if you remember this chart, 13 00:00:40,950 --> 00:00:45,370 we can see that the seasonality of sunspots is far greater than two years. 14 00:00:45,370 --> 00:00:47,262 It's closer to 11 years. 15 00:00:47,262 --> 00:00:51,634 And actually some science tells us that it might even be 22 years with different 16 00:00:51,634 --> 00:00:53,833 cycles interleaguing with each other. 17 00:00:53,833 --> 00:00:58,179 So what would happen if we retrain with a window size of 132, 18 00:00:58,179 --> 00:01:01,790 which is 11 years worth of data as our window size. 19 00:01:03,060 --> 00:01:07,510 Now while this chart looks similar, we can see from the MAE that it actually got 20 00:01:07,510 --> 00:01:11,320 worse so increasing the window size didn't work. 21 00:01:11,320 --> 00:01:12,280 Why do you think that would be? 22 00:01:13,570 --> 00:01:17,630 Well, by looking back to the data, we can realize that it is seasonal to 23 00:01:17,630 --> 00:01:22,530 about 11 years, but we don't need a full season in our window. 24 00:01:22,530 --> 00:01:24,270 Zooming in on the data again, 25 00:01:24,270 --> 00:01:29,250 we'll see something like this where it's just the typical time series. 26 00:01:29,250 --> 00:01:34,310 Values later on are somewhat related to earlier ones, but that's a lot of noise. 27 00:01:34,310 --> 00:01:37,890 So maybe we don't need a huge window of time in order to train. 28 00:01:37,890 --> 00:01:40,620 Maybe we should go with something a little bit more like our initial 20, 29 00:01:40,620 --> 00:01:41,829 let's try 30. 30 00:01:43,590 --> 00:01:47,600 So if we look back at this code, we can change our window size to 30. 31 00:01:47,600 --> 00:01:52,999 But then look at the split time, the data set has around 3,500 items of data, 32 00:01:52,999 --> 00:01:56,689 but we're splitting it into training and validation. 33 00:01:56,689 --> 00:02:03,080 Now 1,000, which means only 1,000 for training and 2,500 for validation. 34 00:02:03,080 --> 00:02:04,462 That's a really bad split. 35 00:02:04,462 --> 00:02:06,431 There's not enough training data. 36 00:02:06,431 --> 00:02:09,860 So let's make it 3,500 instead. 37 00:02:09,860 --> 00:02:13,320 And then when we retrain, we'll get this. 38 00:02:13,320 --> 00:02:17,947 Our MAE has improved to 15 but can we make it even better? 39 00:02:17,947 --> 00:02:21,491 Well, one thing we can try is to edit the neural network design and 40 00:02:21,491 --> 00:02:22,786 height of parameters. 41 00:02:22,786 --> 00:02:25,540 If you remember, we had three layers of 10, 10, and 1 neurons. 42 00:02:25,540 --> 00:02:29,840 Our input shape is now larger at 30. 43 00:02:29,840 --> 00:02:35,170 So maybe try different values here, like 30, 15, and 1, and retrain. 44 00:02:36,410 --> 00:02:41,030 Surprisingly, this was a small step backwards, with our MAE increasing. 45 00:02:42,120 --> 00:02:46,690 It also wasn't worth the extra compute time for the extra neuron layers. 46 00:02:46,690 --> 00:02:50,570 So let's switch back to 10, 10, 1 and instead look at the learning rate. 47 00:02:52,080 --> 00:02:52,920 Let's tweak it a little. 48 00:02:54,650 --> 00:02:59,270 Now after retraining, I can see my MAE has decreased a bit which is good.