1
00:00:12,639 --> 00:00:16,240
hello and welcome again to the data

2
00:00:14,639 --> 00:00:19,039
science and analytics track here at

3
00:00:16,240 --> 00:00:21,359
pycon iu 2021

4
00:00:19,039 --> 00:00:23,519
next up from mysore india

5
00:00:21,359 --> 00:00:25,039
is gaziandra deshpande

6
00:00:23,519 --> 00:00:27,359
they held a masters in technology in

7
00:00:25,039 --> 00:00:28,800
computer science and engineering from

8
00:00:27,359 --> 00:00:31,279
um

9
00:00:28,800 --> 00:00:33,120
oh damn i got the thing wrong

10
00:00:31,279 --> 00:00:34,800
and also a postgraduate diploma in cyber

11
00:00:33,120 --> 00:00:36,239
law and cyber forensics from national

12
00:00:34,800 --> 00:00:37,520
law school of india university in

13
00:00:36,239 --> 00:00:39,360
bengaluru

14
00:00:37,520 --> 00:00:41,440
they have presented talks and papers and

15
00:00:39,360 --> 00:00:43,200
been reviewer and program committee at

16
00:00:41,440 --> 00:00:44,480
more conference and more continents that

17
00:00:43,200 --> 00:00:47,760
i can count

18
00:00:44,480 --> 00:00:49,760
and gazendra is also a github certified

19
00:00:47,760 --> 00:00:52,239
campus advisor and they lead the paidada

20
00:00:49,760 --> 00:00:53,920
bellagavi chapter as well as the owaspe

21
00:00:52,239 --> 00:00:56,719
agave chapter

22
00:00:53,920 --> 00:00:58,879
in the talk today gajendra reminds us

23
00:00:56,719 --> 00:01:00,320
that time series is one of the important

24
00:00:58,879 --> 00:01:01,600
as well as lesser known fields of

25
00:01:00,320 --> 00:01:03,520
machine learning

26
00:01:01,600 --> 00:01:05,680
in time series forecasting we analyze

27
00:01:03,520 --> 00:01:06,880
the time dependent data for predicting

28
00:01:05,680 --> 00:01:09,520
long-term

29
00:01:06,880 --> 00:01:11,439
and seasonally trends

30
00:01:09,520 --> 00:01:13,360
please give a warm welcome to gazendra

31
00:01:11,439 --> 00:01:17,200
deshpande and the presentation on time

32
00:01:13,360 --> 00:01:17,200
series forecasting techniques in python

33
00:01:20,799 --> 00:01:24,560
thank you for the nice introduction it

34
00:01:22,799 --> 00:01:28,880
is wonderful to

35
00:01:24,560 --> 00:01:31,759
speak at pycon conferences always

36
00:01:28,880 --> 00:01:33,520
this is my second talk at pycon

37
00:01:31,759 --> 00:01:34,880
australia online

38
00:01:33,520 --> 00:01:37,119
so today i will be presenting a talk on

39
00:01:34,880 --> 00:01:40,400
time series for casting using

40
00:01:37,119 --> 00:01:42,640
python so it will be a basic talk

41
00:01:40,400 --> 00:01:44,320
so we will not be covering uh any

42
00:01:42,640 --> 00:01:46,720
advanced topics here

43
00:01:44,320 --> 00:01:48,799
so in my talk today we will today i will

44
00:01:46,720 --> 00:01:51,119
discuss introduction to time series

45
00:01:48,799 --> 00:01:52,880
forecasting techniques then basic steps

46
00:01:51,119 --> 00:01:55,680
in forecasting

47
00:01:52,880 --> 00:01:57,840
then time series forecasting techniques

48
00:01:55,680 --> 00:02:00,320
then python libraries for

49
00:01:57,840 --> 00:02:01,600
time says forecasting and finally we'll

50
00:02:00,320 --> 00:02:04,079
summarize and we'll take up some

51
00:02:01,600 --> 00:02:04,079
questions

52
00:02:04,240 --> 00:02:08,560
so

53
00:02:05,600 --> 00:02:09,840
we need to first define what is

54
00:02:08,560 --> 00:02:12,879
time series

55
00:02:09,840 --> 00:02:14,720
so it's a collection of data points at

56
00:02:12,879 --> 00:02:18,400
constant time intervals

57
00:02:14,720 --> 00:02:21,360
analyzed to determine a long term trend

58
00:02:18,400 --> 00:02:23,440
so you can see here on this slide

59
00:02:21,360 --> 00:02:25,599
there is a picture which is

60
00:02:23,440 --> 00:02:27,760
showing the temperature data so

61
00:02:25,599 --> 00:02:28,959
temperature collected at

62
00:02:27,760 --> 00:02:29,840
various

63
00:02:28,959 --> 00:02:32,720
hours

64
00:02:29,840 --> 00:02:34,959
and based on the historical data

65
00:02:32,720 --> 00:02:38,800
we are going to predict the future

66
00:02:34,959 --> 00:02:42,560
trends say for example we have collected

67
00:02:38,800 --> 00:02:43,840
uh temperature data for say around

68
00:02:42,560 --> 00:02:46,319
one month

69
00:02:43,840 --> 00:02:48,400
okay then we are going to predict what

70
00:02:46,319 --> 00:02:49,519
will be the temperature

71
00:02:48,400 --> 00:02:51,840
next day

72
00:02:49,519 --> 00:02:54,560
or maybe we are

73
00:02:51,840 --> 00:02:56,800
collecting the temperature data

74
00:02:54,560 --> 00:02:59,519
say for the month of september over the

75
00:02:56,800 --> 00:03:01,200
years and we are going to predict what

76
00:02:59,519 --> 00:03:03,440
will be the temperature temperature and

77
00:03:01,200 --> 00:03:05,840
what will be the weather

78
00:03:03,440 --> 00:03:07,760
for september 2022

79
00:03:05,840 --> 00:03:09,920
so this is what we are going to do in

80
00:03:07,760 --> 00:03:11,920
time series forecasting of course it is

81
00:03:09,920 --> 00:03:13,920
not only related to

82
00:03:11,920 --> 00:03:17,280
weather forecasting anything related to

83
00:03:13,920 --> 00:03:19,599
forecasting and which is time dependent

84
00:03:17,280 --> 00:03:23,360
we are going to forecast

85
00:03:19,599 --> 00:03:26,239
and python has a has got great support

86
00:03:23,360 --> 00:03:28,080
for these libraries

87
00:03:26,239 --> 00:03:31,920
now there are some application areas

88
00:03:28,080 --> 00:03:31,920
which i have uh listed

89
00:03:32,080 --> 00:03:36,720
i may not be able to explain everything

90
00:03:35,599 --> 00:03:38,959
but we will

91
00:03:36,720 --> 00:03:40,560
but i will explain few

92
00:03:38,959 --> 00:03:42,480
say for example

93
00:03:40,560 --> 00:03:44,799
weather forecasting i have just

94
00:03:42,480 --> 00:03:48,080
explained so it may

95
00:03:44,799 --> 00:03:50,400
be used in agricultural industry may be

96
00:03:48,080 --> 00:03:51,519
used by farmers also

97
00:03:50,400 --> 00:03:53,280
so

98
00:03:51,519 --> 00:03:54,720
before

99
00:03:53,280 --> 00:03:58,000
sowing these seeds

100
00:03:54,720 --> 00:03:59,840
a farmer wants to know what will be the

101
00:03:58,000 --> 00:04:01,200
weather forecast for

102
00:03:59,840 --> 00:04:03,680
next few days

103
00:04:01,200 --> 00:04:06,080
so that accordingly he can

104
00:04:03,680 --> 00:04:10,560
plan his farming activities

105
00:04:06,080 --> 00:04:10,560
then similarly earthquake prediction

106
00:04:11,040 --> 00:04:13,120
so

107
00:04:11,760 --> 00:04:15,040
there will be

108
00:04:13,120 --> 00:04:17,200
patterns of earthquakes so what will be

109
00:04:15,040 --> 00:04:17,200
the

110
00:04:17,280 --> 00:04:22,079
area whatever the environment conditions

111
00:04:20,079 --> 00:04:22,960
are earthquakes are happening in that

112
00:04:22,079 --> 00:04:25,120
area

113
00:04:22,960 --> 00:04:27,440
frequently so based on that they try to

114
00:04:25,120 --> 00:04:29,199
predict when the earth earthquake may

115
00:04:27,440 --> 00:04:31,759
happen

116
00:04:29,199 --> 00:04:34,160
then statistics

117
00:04:31,759 --> 00:04:36,160
at least has got lots of

118
00:04:34,160 --> 00:04:37,520
libraries and lots of applications

119
00:04:36,160 --> 00:04:40,240
related to

120
00:04:37,520 --> 00:04:40,240
forecasting

121
00:04:40,840 --> 00:04:47,040
then also in pattern recognition

122
00:04:45,520 --> 00:04:48,400
then signal processing control

123
00:04:47,040 --> 00:04:52,240
engineering

124
00:04:48,400 --> 00:04:53,120
then mathematical finance astronomy

125
00:04:52,240 --> 00:04:55,440
so

126
00:04:53,120 --> 00:04:57,280
many applications are there

127
00:04:55,440 --> 00:05:00,479
maybe if i have to give

128
00:04:57,280 --> 00:05:03,280
one more example then

129
00:05:00,479 --> 00:05:04,880
say for example there's a website and i

130
00:05:03,280 --> 00:05:06,080
want to predict the

131
00:05:04,880 --> 00:05:08,000
number of

132
00:05:06,080 --> 00:05:10,720
visitors in

133
00:05:08,000 --> 00:05:12,240
next few days based on the previous data

134
00:05:10,720 --> 00:05:14,560
so historical data is very very

135
00:05:12,240 --> 00:05:16,639
important for us so based on that only

136
00:05:14,560 --> 00:05:19,199
we have to

137
00:05:16,639 --> 00:05:19,199
forecast

138
00:05:19,680 --> 00:05:24,479
now there are

139
00:05:21,039 --> 00:05:25,759
common time series patterns so whenever

140
00:05:24,479 --> 00:05:28,880
we use

141
00:05:25,759 --> 00:05:30,240
time series it is very important to

142
00:05:28,880 --> 00:05:32,320
go for

143
00:05:30,240 --> 00:05:34,639
visualization of the data so that

144
00:05:32,320 --> 00:05:35,680
visualization of the data

145
00:05:34,639 --> 00:05:37,360
will

146
00:05:35,680 --> 00:05:39,120
give us

147
00:05:37,360 --> 00:05:41,520
uh

148
00:05:39,120 --> 00:05:42,560
the patterns in the

149
00:05:41,520 --> 00:05:43,440
data

150
00:05:42,560 --> 00:05:45,680
okay

151
00:05:43,440 --> 00:05:45,680
so

152
00:05:45,759 --> 00:05:49,680
you can see here there are five

153
00:05:48,400 --> 00:05:51,680
time series

154
00:05:49,680 --> 00:05:55,120
common time series patterns

155
00:05:51,680 --> 00:05:56,720
so first one is level then second one is

156
00:05:55,120 --> 00:05:58,960
trend

157
00:05:56,720 --> 00:05:58,960
then

158
00:05:59,120 --> 00:06:02,639
next one is um

159
00:06:00,639 --> 00:06:04,840
seasonal or periodic

160
00:06:02,639 --> 00:06:06,400
then cyclical then

161
00:06:04,840 --> 00:06:08,400
irregular

162
00:06:06,400 --> 00:06:11,280
fluctuations

163
00:06:08,400 --> 00:06:13,680
so first one is irregular fluctuations

164
00:06:11,280 --> 00:06:15,840
irregular fluctuations are nothing but

165
00:06:13,680 --> 00:06:19,199
they are random in nature

166
00:06:15,840 --> 00:06:22,000
and most of the times this data will be

167
00:06:19,199 --> 00:06:24,080
random so we can't say that the data

168
00:06:22,000 --> 00:06:25,919
will always going to increase our data

169
00:06:24,080 --> 00:06:28,080
will always going to decrease of course

170
00:06:25,919 --> 00:06:29,360
there will be a pattern

171
00:06:28,080 --> 00:06:32,080
in that

172
00:06:29,360 --> 00:06:33,039
but irregular fluctuations is a random

173
00:06:32,080 --> 00:06:35,680
data

174
00:06:33,039 --> 00:06:37,199
there is a lot of randomness

175
00:06:35,680 --> 00:06:39,840
in in data

176
00:06:37,199 --> 00:06:40,960
the next one is the

177
00:06:39,840 --> 00:06:42,479
trend

178
00:06:40,960 --> 00:06:45,440
so trend

179
00:06:42,479 --> 00:06:47,840
shows whether the data is increasing or

180
00:06:45,440 --> 00:06:51,199
decreasing so on this slide you can see

181
00:06:47,840 --> 00:06:52,720
that there is a line here

182
00:06:51,199 --> 00:06:54,800
so which is

183
00:06:52,720 --> 00:06:57,360
growing so that means the trend is

184
00:06:54,800 --> 00:06:58,880
increasing of course trend may be

185
00:06:57,360 --> 00:07:02,400
decreasing also

186
00:06:58,880 --> 00:07:07,599
but in this example it is not shown

187
00:07:02,400 --> 00:07:07,599
then next is the seasonal or periodic

188
00:07:09,120 --> 00:07:13,840
say for example in india we have got

189
00:07:11,360 --> 00:07:15,520
three seasons one is rainy season and

190
00:07:13,840 --> 00:07:16,800
second one is winter season and third

191
00:07:15,520 --> 00:07:18,319
one is

192
00:07:16,800 --> 00:07:20,160
summer season

193
00:07:18,319 --> 00:07:23,280
so there's a lot of temperature

194
00:07:20,160 --> 00:07:23,280
variation and there is

195
00:07:23,919 --> 00:07:29,120
say forever in summer it will be a hot

196
00:07:26,080 --> 00:07:30,720
season then in

197
00:07:29,120 --> 00:07:32,240
rainy season there will be a lot of rain

198
00:07:30,720 --> 00:07:34,720
and in winter season there will be lot

199
00:07:32,240 --> 00:07:34,720
of cold

200
00:07:35,039 --> 00:07:37,599
so

201
00:07:35,919 --> 00:07:39,759
there will be different temperature in

202
00:07:37,599 --> 00:07:42,080
these seasons

203
00:07:39,759 --> 00:07:42,840
then similarly

204
00:07:42,080 --> 00:07:45,199
these

205
00:07:42,840 --> 00:07:47,199
seasons there are different periods say

206
00:07:45,199 --> 00:07:49,599
for example some pattern may repeat

207
00:07:47,199 --> 00:07:51,759
daily some pattern may repeat weekly

208
00:07:49,599 --> 00:07:55,440
some pattern may repeat monthly

209
00:07:51,759 --> 00:07:55,440
and some pattern may repeat

210
00:07:56,160 --> 00:07:58,479
yearly

211
00:07:57,599 --> 00:08:01,039
okay

212
00:07:58,479 --> 00:08:03,680
in astronomy in astronomy also there are

213
00:08:01,039 --> 00:08:05,919
some patterns say for example

214
00:08:03,680 --> 00:08:05,919
uh

215
00:08:06,720 --> 00:08:11,039
uh maybe

216
00:08:08,160 --> 00:08:12,800
the moving cycles or the

217
00:08:11,039 --> 00:08:14,960
uh

218
00:08:12,800 --> 00:08:18,000
the constellations

219
00:08:14,960 --> 00:08:19,840
they have got some pattern of occurrence

220
00:08:18,000 --> 00:08:22,400
the next is the

221
00:08:19,840 --> 00:08:24,000
cyclical so in cyclical

222
00:08:22,400 --> 00:08:27,120
we are going to have

223
00:08:24,000 --> 00:08:28,400
both patterns increasing as well as

224
00:08:27,120 --> 00:08:30,560
decreasing

225
00:08:28,400 --> 00:08:32,880
but it will not be a just

226
00:08:30,560 --> 00:08:34,399
one phase there will be multiple

227
00:08:32,880 --> 00:08:35,519
repetitive

228
00:08:34,399 --> 00:08:39,279
phases

229
00:08:35,519 --> 00:08:41,519
so you can see here that there is a

230
00:08:39,279 --> 00:08:43,360
upward trend then downward and then

231
00:08:41,519 --> 00:08:46,560
again upward trend so this is cyclical

232
00:08:43,360 --> 00:08:49,200
then again uh downward trend so this is

233
00:08:46,560 --> 00:08:51,360
cyclical so irregular fluctuations we

234
00:08:49,200 --> 00:08:52,880
have discussed that it's a

235
00:08:51,360 --> 00:08:54,800
random one

236
00:08:52,880 --> 00:08:57,360
so these are the common time series

237
00:08:54,800 --> 00:08:58,640
patterns and of course this also affects

238
00:08:57,360 --> 00:09:01,440
our

239
00:08:58,640 --> 00:09:01,440
our results

240
00:09:04,000 --> 00:09:09,760
then there are some basic steps in

241
00:09:06,839 --> 00:09:12,160
performing forecasting

242
00:09:09,760 --> 00:09:14,800
so first one is the defining a problem

243
00:09:12,160 --> 00:09:16,800
then collect the information

244
00:09:14,800 --> 00:09:18,800
then perform preliminary

245
00:09:16,800 --> 00:09:20,959
explorative analysis

246
00:09:18,800 --> 00:09:24,800
then choose and

247
00:09:20,959 --> 00:09:26,480
fit model then use and evaluate

248
00:09:24,800 --> 00:09:27,440
a forecasting

249
00:09:26,480 --> 00:09:28,720
model

250
00:09:27,440 --> 00:09:30,640
so first one

251
00:09:28,720 --> 00:09:32,399
is the defining a problem which is very

252
00:09:30,640 --> 00:09:35,600
very important so

253
00:09:32,399 --> 00:09:37,360
we need a problem to work on

254
00:09:35,600 --> 00:09:39,120
we need a date

255
00:09:37,360 --> 00:09:40,720
we need a problem to work on and here we

256
00:09:39,120 --> 00:09:42,000
need to define what is our problem and

257
00:09:40,720 --> 00:09:43,120
what we want to

258
00:09:42,000 --> 00:09:45,760
achieve

259
00:09:43,120 --> 00:09:47,760
so problem and goal should be defined

260
00:09:45,760 --> 00:09:50,000
here the next is

261
00:09:47,760 --> 00:09:51,680
collect the information so here we are

262
00:09:50,000 --> 00:09:53,839
building the

263
00:09:51,680 --> 00:09:55,440
data set by collecting the

264
00:09:53,839 --> 00:09:57,600
historical data

265
00:09:55,440 --> 00:10:00,720
so we also need

266
00:09:57,600 --> 00:10:02,560
inputs from the domain experts

267
00:10:00,720 --> 00:10:05,839
if your domain expert you can directly

268
00:10:02,560 --> 00:10:07,519
work on it then perform a preliminary

269
00:10:05,839 --> 00:10:08,720
explorative

270
00:10:07,519 --> 00:10:10,800
analysis

271
00:10:08,720 --> 00:10:13,920
so in preliminary explorative analysis

272
00:10:10,800 --> 00:10:16,000
you are going to use tools like um

273
00:10:13,920 --> 00:10:19,800
visualization tools

274
00:10:16,000 --> 00:10:19,800
to identify the

275
00:10:20,240 --> 00:10:23,680
kind of

276
00:10:21,839 --> 00:10:24,800
data pattern

277
00:10:23,680 --> 00:10:26,560
and

278
00:10:24,800 --> 00:10:27,760
once you identify this data pattern then

279
00:10:26,560 --> 00:10:29,440
you can

280
00:10:27,760 --> 00:10:30,480
select a model to

281
00:10:29,440 --> 00:10:32,800
fit

282
00:10:30,480 --> 00:10:34,560
okay so in building a preliminary

283
00:10:32,800 --> 00:10:36,640
explorative analysis you are basically

284
00:10:34,560 --> 00:10:39,760
trying to find out the

285
00:10:36,640 --> 00:10:41,680
pattern in the data by performing

286
00:10:39,760 --> 00:10:44,160
several operations

287
00:10:41,680 --> 00:10:48,160
such as visualization and so on

288
00:10:44,160 --> 00:10:50,320
then choose and fit a model

289
00:10:48,160 --> 00:10:52,959
here you are going to experiment with

290
00:10:50,320 --> 00:10:54,079
different models so the one which gives

291
00:10:52,959 --> 00:10:55,839
you

292
00:10:54,079 --> 00:10:57,680
less error

293
00:10:55,839 --> 00:11:01,839
compared to the other

294
00:10:57,680 --> 00:11:01,839
model you are going to select that

295
00:11:02,480 --> 00:11:07,839
then use and evaluate

296
00:11:05,680 --> 00:11:09,440
forecasting model in your application so

297
00:11:07,839 --> 00:11:10,560
finally you are going to

298
00:11:09,440 --> 00:11:14,120
deploy it

299
00:11:10,560 --> 00:11:14,120
in your application

300
00:11:16,079 --> 00:11:19,760
now these are the

301
00:11:18,160 --> 00:11:22,880
common time series

302
00:11:19,760 --> 00:11:25,040
methods uh supported by stats models

303
00:11:22,880 --> 00:11:27,360
package in python

304
00:11:25,040 --> 00:11:28,880
so first one is the auto regression then

305
00:11:27,360 --> 00:11:30,399
moving average

306
00:11:28,880 --> 00:11:32,399
so these are the

307
00:11:30,399 --> 00:11:34,399
two basic

308
00:11:32,399 --> 00:11:35,680
time series methods

309
00:11:34,399 --> 00:11:36,880
but

310
00:11:35,680 --> 00:11:37,839
if you see the

311
00:11:36,880 --> 00:11:39,279
rest

312
00:11:37,839 --> 00:11:41,839
then

313
00:11:39,279 --> 00:11:44,880
they are the combinations of first two

314
00:11:41,839 --> 00:11:44,880
and apart from that

315
00:11:45,279 --> 00:11:49,760
the

316
00:11:47,200 --> 00:11:53,200
different parameters such as seasonality

317
00:11:49,760 --> 00:11:54,720
has been added to the

318
00:11:53,200 --> 00:11:55,519
basic methods

319
00:11:54,720 --> 00:11:57,760
okay

320
00:11:55,519 --> 00:11:59,839
so first one is auto regression then

321
00:11:57,760 --> 00:12:02,000
moving average then auto regressive

322
00:11:59,839 --> 00:12:04,000
moving average auto regressive

323
00:12:02,000 --> 00:12:06,160
integrated

324
00:12:04,000 --> 00:12:08,639
moving average then seasonal auto

325
00:12:06,160 --> 00:12:10,560
regressive integrated moving average

326
00:12:08,639 --> 00:12:13,680
then seasonal autoregressive integrated

327
00:12:10,560 --> 00:12:15,680
moving average with exogenous regressors

328
00:12:13,680 --> 00:12:16,720
and vector auto regression moving

329
00:12:15,680 --> 00:12:18,399
average

330
00:12:16,720 --> 00:12:21,680
then hold winston's exponential

331
00:12:18,399 --> 00:12:24,680
smoothing and finally the diki fuller

332
00:12:21,680 --> 00:12:24,680
test

333
00:12:26,160 --> 00:12:31,040
what happens in auto regression is that

334
00:12:28,320 --> 00:12:35,040
we are going to predict the

335
00:12:31,040 --> 00:12:38,639
future based on the past data

336
00:12:35,040 --> 00:12:40,800
so this past data is given as input to

337
00:12:38,639 --> 00:12:42,560
the regression

338
00:12:40,800 --> 00:12:45,040
then it will be

339
00:12:42,560 --> 00:12:46,639
then it will try to predict the

340
00:12:45,040 --> 00:12:47,839
uh

341
00:12:46,639 --> 00:12:50,240
values

342
00:12:47,839 --> 00:12:52,399
okay so we can just

343
00:12:50,240 --> 00:12:55,399
try to test this

344
00:12:52,399 --> 00:12:55,399
application

345
00:13:17,279 --> 00:13:21,360
okay so before that

346
00:13:19,600 --> 00:13:23,600
so let's see the source code so what we

347
00:13:21,360 --> 00:13:25,200
are doing here is we are trying to

348
00:13:23,600 --> 00:13:28,000
import the

349
00:13:25,200 --> 00:13:29,839
uh package

350
00:13:28,000 --> 00:13:30,880
import the modulator right then we are

351
00:13:29,839 --> 00:13:32,959
also

352
00:13:30,880 --> 00:13:35,680
importing the random module because we

353
00:13:32,959 --> 00:13:37,760
want to generate the data set random

354
00:13:35,680 --> 00:13:39,040
this is just for

355
00:13:37,760 --> 00:13:42,880
an example

356
00:13:39,040 --> 00:13:45,440
but you can really go for real-world

357
00:13:42,880 --> 00:13:45,440
data set

358
00:13:45,519 --> 00:13:49,360
then

359
00:13:46,560 --> 00:13:51,120
fit the model

360
00:13:49,360 --> 00:13:52,639
so here we are using the

361
00:13:51,120 --> 00:13:55,279
autoric function

362
00:13:52,639 --> 00:13:57,519
and it takes

363
00:13:55,279 --> 00:14:00,720
two parameters data and lags

364
00:13:57,519 --> 00:14:03,440
so here lags is the

365
00:14:00,720 --> 00:14:06,480
pattern so here we are predict we want

366
00:14:03,440 --> 00:14:06,480
to put the one value

367
00:14:07,519 --> 00:14:11,600
then fit the model

368
00:14:09,360 --> 00:14:12,880
then make the prediction

369
00:14:11,600 --> 00:14:15,880
and print the

370
00:14:12,880 --> 00:14:15,880
prediction

371
00:14:18,320 --> 00:14:22,639
okay so next is uh

372
00:14:20,720 --> 00:14:24,720
the moving average

373
00:14:22,639 --> 00:14:27,279
so in moving average what happens is we

374
00:14:24,720 --> 00:14:27,279
are going to

375
00:14:28,839 --> 00:14:35,760
uh the average so based on

376
00:14:32,959 --> 00:14:38,720
selected data points

377
00:14:35,760 --> 00:14:40,240
say last n data points are taken to

378
00:14:38,720 --> 00:14:42,160
perform the

379
00:14:40,240 --> 00:14:43,920
moving average

380
00:14:42,160 --> 00:14:45,360
or to calculate the to calculate the

381
00:14:43,920 --> 00:14:49,120
moving average

382
00:14:45,360 --> 00:14:50,639
so here we are importing the

383
00:14:49,120 --> 00:14:52,800
arima model

384
00:14:50,639 --> 00:14:52,800
uh

385
00:14:53,680 --> 00:15:00,320
because we don't have any direct

386
00:14:56,320 --> 00:15:00,320
moving average module here

387
00:15:01,760 --> 00:15:05,760
but the first parameter is

388
00:15:04,079 --> 00:15:09,440
set as uh

389
00:15:05,760 --> 00:15:10,839
zero so that is going to make it as

390
00:15:09,440 --> 00:15:13,199
moving

391
00:15:10,839 --> 00:15:15,519
average so

392
00:15:13,199 --> 00:15:17,920
we are making first parameters in order

393
00:15:15,519 --> 00:15:20,399
that is auto regression and integrated

394
00:15:17,920 --> 00:15:20,399
as zero

395
00:15:20,720 --> 00:15:24,240
so

396
00:15:22,160 --> 00:15:27,839
it will compute the moving average on

397
00:15:24,240 --> 00:15:27,839
the random data

398
00:15:31,759 --> 00:15:34,959
the next is auto regressive moving

399
00:15:34,000 --> 00:15:37,279
average

400
00:15:34,959 --> 00:15:39,680
it is the combination of

401
00:15:37,279 --> 00:15:41,759
of previous two

402
00:15:39,680 --> 00:15:43,519
models

403
00:15:41,759 --> 00:15:46,160
that is the

404
00:15:43,519 --> 00:15:48,959
auto regress auto regression and

405
00:15:46,160 --> 00:15:48,959
moving average

406
00:15:51,839 --> 00:15:57,199
and third one is

407
00:15:55,120 --> 00:15:59,360
auto regressive integrated moving

408
00:15:57,199 --> 00:16:01,600
average arima again it's a combination

409
00:15:59,360 --> 00:16:05,120
of previous two but here the

410
00:16:01,600 --> 00:16:07,519
additional parameter has been added

411
00:16:05,120 --> 00:16:10,000
so that is the

412
00:16:07,519 --> 00:16:12,079
integrating feature

413
00:16:10,000 --> 00:16:14,800
so the integrating feature

414
00:16:12,079 --> 00:16:16,639
is related to the

415
00:16:14,800 --> 00:16:17,920
seasonality

416
00:16:16,639 --> 00:16:20,240
trends

417
00:16:17,920 --> 00:16:20,240
here

418
00:16:21,199 --> 00:16:24,560
uh not seasonality sorry it's it's

419
00:16:23,199 --> 00:16:26,240
related to the

420
00:16:24,560 --> 00:16:29,440
stationarity

421
00:16:26,240 --> 00:16:32,079
so what happens is um

422
00:16:29,440 --> 00:16:33,839
if your data is not stationary then you

423
00:16:32,079 --> 00:16:35,839
have to make it stationary whenever you

424
00:16:33,839 --> 00:16:37,199
are performing

425
00:16:35,839 --> 00:16:39,600
time series

426
00:16:37,199 --> 00:16:44,240
forecasting

427
00:16:39,600 --> 00:16:44,240
so that is what happens in arima model

428
00:16:46,959 --> 00:16:52,480
then in

429
00:16:48,000 --> 00:16:55,120
sarima model the seasonality values are

430
00:16:52,480 --> 00:16:59,040
are added

431
00:16:55,120 --> 00:17:01,680
so it directly supports the

432
00:16:59,040 --> 00:17:04,480
it directly combines the arima model

433
00:17:01,680 --> 00:17:07,120
with the ability of again auto

434
00:17:04,480 --> 00:17:10,640
regression then

435
00:17:07,120 --> 00:17:13,360
differencing then moving average

436
00:17:10,640 --> 00:17:16,640
and at the seasonal level

437
00:17:13,360 --> 00:17:19,120
so it is suitable for again univariate

438
00:17:16,640 --> 00:17:21,919
time series

439
00:17:19,120 --> 00:17:23,919
or trend or seasonal components so again

440
00:17:21,919 --> 00:17:26,480
all previous methods whatever we have

441
00:17:23,919 --> 00:17:29,760
discussed they are all related to or

442
00:17:26,480 --> 00:17:29,760
they are all suitable for

443
00:17:29,840 --> 00:17:32,840
uh

444
00:17:35,919 --> 00:17:39,760
they are also double for univariate time

445
00:17:38,080 --> 00:17:41,039
series analysis so there are some

446
00:17:39,760 --> 00:17:43,120
methods which are suitable for

447
00:17:41,039 --> 00:17:45,120
multivariate analysis which

448
00:17:43,120 --> 00:17:47,760
we will see later

449
00:17:45,120 --> 00:17:51,600
now all these methods in

450
00:17:47,760 --> 00:17:53,360
inside they use a kalman filter

451
00:17:51,600 --> 00:17:54,960
for

452
00:17:53,360 --> 00:17:56,400
fitting the values

453
00:17:54,960 --> 00:17:59,400
using the predict

454
00:17:56,400 --> 00:17:59,400
method

455
00:18:02,480 --> 00:18:08,799
now next one is

456
00:18:05,520 --> 00:18:10,320
cerimo model with exogenous

457
00:18:08,799 --> 00:18:13,520
regressors

458
00:18:10,320 --> 00:18:15,280
now here you can see here that

459
00:18:13,520 --> 00:18:16,559
we are using two

460
00:18:15,280 --> 00:18:18,960
data sets

461
00:18:16,559 --> 00:18:21,600
so data one and

462
00:18:18,960 --> 00:18:21,600
data two

463
00:18:23,360 --> 00:18:25,520
so

464
00:18:26,400 --> 00:18:30,480
we are saying that

465
00:18:28,960 --> 00:18:32,799
in ceramics

466
00:18:30,480 --> 00:18:34,880
function the first parameter is the

467
00:18:32,799 --> 00:18:37,919
data and second parameter is the

468
00:18:34,880 --> 00:18:39,360
exogenous regressor

469
00:18:37,919 --> 00:18:40,240
values

470
00:18:39,360 --> 00:18:42,400
so

471
00:18:40,240 --> 00:18:44,640
they are exogenous

472
00:18:42,400 --> 00:18:46,720
variables are also called as the

473
00:18:44,640 --> 00:18:48,080
covariates

474
00:18:46,720 --> 00:18:49,760
and they can be

475
00:18:48,080 --> 00:18:52,320
thought of as a

476
00:18:49,760 --> 00:18:54,799
parallel input sequences

477
00:18:52,320 --> 00:18:57,600
so that have observations in

478
00:18:54,799 --> 00:19:01,600
at the same steps at the original

479
00:18:57,600 --> 00:19:01,600
original series so that means

480
00:19:02,080 --> 00:19:05,280
at the same time

481
00:19:03,840 --> 00:19:07,679
there are

482
00:19:05,280 --> 00:19:10,720
multiple values which are

483
00:19:07,679 --> 00:19:13,760
being observed

484
00:19:10,720 --> 00:19:15,679
so again this method is suitable for

485
00:19:13,760 --> 00:19:18,320
univariate

486
00:19:15,679 --> 00:19:18,320
time series

487
00:19:22,960 --> 00:19:27,600
now vector auto regression moving

488
00:19:25,600 --> 00:19:30,320
average

489
00:19:27,600 --> 00:19:32,400
this uses

490
00:19:30,320 --> 00:19:33,520
auto regression and moving average

491
00:19:32,400 --> 00:19:35,840
methods

492
00:19:33,520 --> 00:19:37,679
combined with

493
00:19:35,840 --> 00:19:39,919
vectors

494
00:19:37,679 --> 00:19:41,840
so this is suitable for this is suitable

495
00:19:39,919 --> 00:19:46,400
for multivariate

496
00:19:41,840 --> 00:19:48,720
time analysis multivariate time series

497
00:19:46,400 --> 00:19:50,720
time series forecasting

498
00:19:48,720 --> 00:19:54,880
so that's why you can see here that in

499
00:19:50,720 --> 00:19:56,799
this code the list is created then

500
00:19:54,880 --> 00:19:59,360
the first vector is

501
00:19:56,799 --> 00:20:02,080
generated randomly where the second

502
00:19:59,360 --> 00:20:03,840
vectors values are dependent on the

503
00:20:02,080 --> 00:20:07,440
first vector

504
00:20:03,840 --> 00:20:09,280
so basically you need here two sets of

505
00:20:07,440 --> 00:20:12,400
values

506
00:20:09,280 --> 00:20:14,559
so a row will be having both the values

507
00:20:12,400 --> 00:20:16,400
vector one as well as vector two now

508
00:20:14,559 --> 00:20:18,840
note here that vector two's value is

509
00:20:16,400 --> 00:20:22,720
always dependent on the vector

510
00:20:18,840 --> 00:20:25,200
one so this is suitable for

511
00:20:22,720 --> 00:20:26,559
multivariate time series

512
00:20:25,200 --> 00:20:28,000
forecasting

513
00:20:26,559 --> 00:20:30,559
but

514
00:20:28,000 --> 00:20:30,559
without

515
00:20:30,799 --> 00:20:36,480
but without

516
00:20:32,240 --> 00:20:36,480
trend and seasonal components

517
00:20:40,960 --> 00:20:45,120
the next one is the holtminster's

518
00:20:43,679 --> 00:20:46,320
exponential

519
00:20:45,120 --> 00:20:49,039
smoothing

520
00:20:46,320 --> 00:20:51,679
so this using

521
00:20:49,039 --> 00:20:54,720
this method uses the

522
00:20:51,679 --> 00:20:54,720
smoothing techniques

523
00:20:55,520 --> 00:21:00,240
but specifically

524
00:20:57,280 --> 00:21:03,200
holtwinster's exponential smoothing uses

525
00:21:00,240 --> 00:21:05,200
triple exponential smoothing method

526
00:21:03,200 --> 00:21:07,600
and it models the

527
00:21:05,200 --> 00:21:10,000
next time step as the exponentially

528
00:21:07,600 --> 00:21:12,000
weighted linear function

529
00:21:10,000 --> 00:21:14,080
of observations

530
00:21:12,000 --> 00:21:16,320
at prior time steps

531
00:21:14,080 --> 00:21:17,120
then taking trends and seasonality into

532
00:21:16,320 --> 00:21:19,520
the

533
00:21:17,120 --> 00:21:19,520
account

534
00:21:20,720 --> 00:21:25,200
so again it is suitable for univariate

535
00:21:23,120 --> 00:21:26,320
time series but with

536
00:21:25,200 --> 00:21:27,520
trend or

537
00:21:26,320 --> 00:21:30,520
seasonal

538
00:21:27,520 --> 00:21:30,520
components

539
00:21:34,720 --> 00:21:37,679
then next is the

540
00:21:38,480 --> 00:21:42,960
package called as

541
00:21:40,559 --> 00:21:45,840
profit it is by

542
00:21:42,960 --> 00:21:45,840
by facebook

543
00:21:47,039 --> 00:21:50,960
so it's a open source software

544
00:21:49,360 --> 00:21:52,799
addressed by facebook

545
00:21:50,960 --> 00:21:55,440
core data science team

546
00:21:52,799 --> 00:21:58,799
so it's a procedure for forecasting

547
00:21:55,440 --> 00:21:59,840
time series data based on an additive

548
00:21:58,799 --> 00:22:02,480
model

549
00:21:59,840 --> 00:22:04,880
where non-linear trends are

550
00:22:02,480 --> 00:22:06,159
fit with yearly weekly and daily

551
00:22:04,880 --> 00:22:08,880
seasonality

552
00:22:06,159 --> 00:22:10,559
plus holiday effects

553
00:22:08,880 --> 00:22:12,720
it works best with

554
00:22:10,559 --> 00:22:13,840
time series that have strong seasonal

555
00:22:12,720 --> 00:22:17,200
effects

556
00:22:13,840 --> 00:22:19,039
and several seasons of historical data

557
00:22:17,200 --> 00:22:22,720
it's a robust

558
00:22:19,039 --> 00:22:25,280
to missing data and shifts

559
00:22:22,720 --> 00:22:26,559
in in the trend and typically handles

560
00:22:25,280 --> 00:22:27,840
outliers

561
00:22:26,559 --> 00:22:30,080
well

562
00:22:27,840 --> 00:22:31,919
so you can visit this url for more

563
00:22:30,080 --> 00:22:34,480
information

564
00:22:31,919 --> 00:22:36,640
it is also it also supports

565
00:22:34,480 --> 00:22:36,640
r

566
00:22:37,520 --> 00:22:41,840
so it has been widely used in

567
00:22:42,000 --> 00:22:46,559
facebook and its

568
00:22:43,760 --> 00:22:46,559
applications

569
00:22:48,400 --> 00:22:52,400
and the next package is the

570
00:22:50,880 --> 00:22:54,720
darts

571
00:22:52,400 --> 00:22:56,799
so it's a python library for

572
00:22:54,720 --> 00:22:59,039
easy manipulation and

573
00:22:56,799 --> 00:23:00,400
forecasting of

574
00:22:59,039 --> 00:23:02,960
time series

575
00:23:00,400 --> 00:23:04,159
so it contains very variety of models

576
00:23:02,960 --> 00:23:06,159
from

577
00:23:04,159 --> 00:23:07,520
classics such as arima to deep neural

578
00:23:06,159 --> 00:23:10,320
networks

579
00:23:07,520 --> 00:23:11,919
the models can all be used in the

580
00:23:10,320 --> 00:23:14,000
same way using

581
00:23:11,919 --> 00:23:15,760
fit and predict functions similar to

582
00:23:14,000 --> 00:23:18,240
scikit lab

583
00:23:15,760 --> 00:23:20,000
then it also makes it

584
00:23:18,240 --> 00:23:22,400
easy to backtest models and combine

585
00:23:20,000 --> 00:23:23,919
predictions of several models

586
00:23:22,400 --> 00:23:25,280
and external

587
00:23:23,919 --> 00:23:27,120
regressors

588
00:23:25,280 --> 00:23:29,840
so it supports both univariate and

589
00:23:27,120 --> 00:23:31,760
multivariate time series and models

590
00:23:29,840 --> 00:23:34,320
then neural networks can be trained on

591
00:23:31,760 --> 00:23:37,919
multiple time series and some

592
00:23:34,320 --> 00:23:40,480
of the models offer the probabilistic

593
00:23:37,919 --> 00:23:43,440
forecasts

594
00:23:40,480 --> 00:23:46,080
now if you see this slide then it shows

595
00:23:43,440 --> 00:23:47,520
the various capabilities of

596
00:23:46,080 --> 00:23:49,039
a darts um

597
00:23:47,520 --> 00:23:52,400
library

598
00:23:49,039 --> 00:23:54,080
so for discovery it supports

599
00:23:52,400 --> 00:23:56,080
seasonality and trend checks

600
00:23:54,080 --> 00:23:59,279
visualizations

601
00:23:56,080 --> 00:24:01,279
then new models

602
00:23:59,279 --> 00:24:04,400
then for preprocessing it supports

603
00:24:01,279 --> 00:24:06,080
normalization then interpolation of

604
00:24:04,400 --> 00:24:09,039
missing values

605
00:24:06,080 --> 00:24:10,400
then seasonality or

606
00:24:09,039 --> 00:24:11,679
or trend

607
00:24:10,400 --> 00:24:12,880
removal

608
00:24:11,679 --> 00:24:15,679
then up

609
00:24:12,880 --> 00:24:18,000
or down sampling

610
00:24:15,679 --> 00:24:20,480
then it supports classic forecasting

611
00:24:18,000 --> 00:24:23,120
models so which we have

612
00:24:20,480 --> 00:24:24,159
seen in our slides

613
00:24:23,120 --> 00:24:26,640
such as

614
00:24:24,159 --> 00:24:29,120
auto regression moving average

615
00:24:26,640 --> 00:24:30,799
arima ceremony max

616
00:24:29,120 --> 00:24:33,679
etc

617
00:24:30,799 --> 00:24:35,600
then it also supports it also integrates

618
00:24:33,679 --> 00:24:37,440
the facebook's

619
00:24:35,600 --> 00:24:38,400
profit library

620
00:24:37,440 --> 00:24:40,080
and

621
00:24:38,400 --> 00:24:42,240
many others

622
00:24:40,080 --> 00:24:44,240
then for model selection and evaluation

623
00:24:42,240 --> 00:24:48,240
it supports back testing

624
00:24:44,240 --> 00:24:51,120
then residual analysis then research

625
00:24:48,240 --> 00:24:53,600
and it also supports various

626
00:24:51,120 --> 00:24:53,600
metrics

627
00:24:54,320 --> 00:24:59,760
so in summary we have discussed about

628
00:24:57,279 --> 00:25:01,440
basics of time series forecasting

629
00:24:59,760 --> 00:25:02,480
then different times is forecasting

630
00:25:01,440 --> 00:25:04,000
techniques

631
00:25:02,480 --> 00:25:05,679
then different python libraries for

632
00:25:04,000 --> 00:25:08,000
performing time series forecasting

633
00:25:05,679 --> 00:25:10,320
quickly of course these are not the only

634
00:25:08,000 --> 00:25:12,080
libraries we have also got other

635
00:25:10,320 --> 00:25:13,840
libraries such as you can also go for

636
00:25:12,080 --> 00:25:17,360
tensorflow so there is a nice

637
00:25:13,840 --> 00:25:17,360
documentation available

638
00:25:18,480 --> 00:25:23,120
so depending on your application

639
00:25:20,400 --> 00:25:24,960
depending on your

640
00:25:23,120 --> 00:25:26,720
depending on the

641
00:25:24,960 --> 00:25:29,039
pattern in the data

642
00:25:26,720 --> 00:25:30,400
you should be selecting the model and

643
00:25:29,039 --> 00:25:33,600
when you are selecting the model of

644
00:25:30,400 --> 00:25:35,760
course there are different options

645
00:25:33,600 --> 00:25:36,960
so you need to select that option which

646
00:25:35,760 --> 00:25:42,080
is

647
00:25:36,960 --> 00:25:43,679
giving you less error as well as

648
00:25:42,080 --> 00:25:45,039
the one which is

649
00:25:43,679 --> 00:25:46,480
performing

650
00:25:45,039 --> 00:25:48,799
better

651
00:25:46,480 --> 00:25:52,080
compared to the others

652
00:25:48,799 --> 00:25:54,159
so thank you everyone for attending my

653
00:25:52,080 --> 00:25:57,400
talk

654
00:25:54,159 --> 00:25:57,400
thank you

655
00:26:07,120 --> 00:26:11,440
thank you kashendra

656
00:26:09,279 --> 00:26:14,159
um we have a couple of questions for you

657
00:26:11,440 --> 00:26:16,000
from from the audience i'm gonna read

658
00:26:14,159 --> 00:26:18,480
them to you the first one is

659
00:26:16,000 --> 00:26:20,720
what advice do you have for someone

660
00:26:18,480 --> 00:26:23,120
starting a time series project for the

661
00:26:20,720 --> 00:26:25,840
first time

662
00:26:23,120 --> 00:26:27,120
okay so if you are starting for

663
00:26:25,840 --> 00:26:29,679
if you are starting

664
00:26:27,120 --> 00:26:31,279
a time series project for the first time

665
00:26:29,679 --> 00:26:33,600
then

666
00:26:31,279 --> 00:26:35,279
first learn about the

667
00:26:33,600 --> 00:26:39,360
statistics and different statistical

668
00:26:35,279 --> 00:26:42,640
methods which we have discussed

669
00:26:39,360 --> 00:26:42,640
then you should have a

670
00:26:43,679 --> 00:26:47,760
basic understanding of

671
00:26:45,520 --> 00:26:51,200
these statistics and mathematics

672
00:26:47,760 --> 00:26:52,640
then learn the basic libraries

673
00:26:51,200 --> 00:26:55,840
in python

674
00:26:52,640 --> 00:26:57,840
so pandas data frames

675
00:26:55,840 --> 00:27:00,480
numpy's scipy

676
00:26:57,840 --> 00:27:02,559
stat models they will help then you can

677
00:27:00,480 --> 00:27:04,000
learn a specific

678
00:27:02,559 --> 00:27:05,679
python package

679
00:27:04,000 --> 00:27:08,240
which is which satisfies your

680
00:27:05,679 --> 00:27:11,679
requirement of course we have got

681
00:27:08,240 --> 00:27:16,000
many options but you should see which

682
00:27:11,679 --> 00:27:16,000
satisfies your requirements better

683
00:27:16,080 --> 00:27:21,120
and then

684
00:27:18,320 --> 00:27:22,720
you have to select a domain

685
00:27:21,120 --> 00:27:25,840
in which you want to

686
00:27:22,720 --> 00:27:27,360
uh perform forecasting so if you have a

687
00:27:25,840 --> 00:27:28,960
domain knowledge then it's fine if you

688
00:27:27,360 --> 00:27:31,679
don't have a domain knowledge then you

689
00:27:28,960 --> 00:27:34,960
need to collaborate with someone

690
00:27:31,679 --> 00:27:37,120
and collect the data and once you have

691
00:27:34,960 --> 00:27:39,520
the data use the packages and you can

692
00:27:37,120 --> 00:27:39,520
start

693
00:27:40,159 --> 00:27:45,440
the forecasting or even start the time

694
00:27:42,559 --> 00:27:45,440
series analysis

695
00:27:46,320 --> 00:27:51,600
um thank you another question they have

696
00:27:48,559 --> 00:27:53,360
is about profit or darts

697
00:27:51,600 --> 00:27:55,200
about performance

698
00:27:53,360 --> 00:27:56,960
do you have any benchmark for the

699
00:27:55,200 --> 00:27:58,559
performance of each

700
00:27:56,960 --> 00:28:01,120
uh library which

701
00:27:58,559 --> 00:28:04,799
um is more performant or better

702
00:28:01,120 --> 00:28:09,360
maybe has a better use in other cases

703
00:28:04,799 --> 00:28:10,720
uh i'm not aware of the uh benchmarks

704
00:28:09,360 --> 00:28:14,080
but yeah maybe

705
00:28:10,720 --> 00:28:18,399
we may use metrics and test

706
00:28:14,080 --> 00:28:18,399
but i'm not aware of the benchmarks

707
00:28:18,960 --> 00:28:25,760
so i think this is it and thank you so

708
00:28:22,559 --> 00:28:28,399
much kajindra and i want to ask the

709
00:28:25,760 --> 00:28:30,960
venulist people to give a big virtual

710
00:28:28,399 --> 00:28:33,840
applause and thank you

711
00:28:30,960 --> 00:28:36,799
yeah thank you

712
00:28:33,840 --> 00:28:38,960
next up on this stage in about 15

713
00:28:36,799 --> 00:28:41,440
minutes at 2 30

714
00:28:38,960 --> 00:28:43,840
um eastern european time we'll have mali

715
00:28:41,440 --> 00:28:46,799
and mangami talking about a beginner's

716
00:28:43,840 --> 00:28:49,799
guide to gpus for pythonistas see you

717
00:28:46,799 --> 00:28:49,799
then