WEBVTT

1
00:00:00.040 --> 00:00:00.240
Hello,

2
00:00:00.420 --> 00:00:02.780
welcome back to Papers with Backtest podcast.

3
00:00:03.040 --> 00:00:05.980
Today we dive into another algo trading research paper.

4
00:00:06.180 --> 00:00:06.880
We do indeed.

5
00:00:07.460 --> 00:00:14.120
We're going to unpack deep learning for forecasting stock returns in the cross-section by Abe and Nakayama from 2018.

6
00:00:14.300 --> 00:00:15.260
A really interesting one.

7
00:00:15.340 --> 00:00:15.520
Yeah,

8
00:00:15.680 --> 00:00:18.200
this paper looks into whether using deep learning can,

9
00:00:18.480 --> 00:00:18.720
you know,

10
00:00:18.760 --> 00:00:22.200
give us a better edge predicting Japanese stock performance over the next month.

11
00:00:22.380 --> 00:00:22.760
Exactly.

12
00:00:22.800 --> 00:00:24.180
It's fundamentally asking,

13
00:00:24.680 --> 00:00:27.340
can these newer,

14
00:00:27.520 --> 00:00:29.660
more complex machine learning tools actually

15
00:00:29.848 --> 00:00:31.089
beat the traditional methods,

16
00:00:31.529 --> 00:00:33.191
especially for forecasting returns.

17
00:00:33.571 --> 00:00:33.831
Right.

18
00:00:33.911 --> 00:00:34.832
So for you listening,

19
00:00:34.872 --> 00:00:41.917
the goal today is figuring out if this deep learning approach really is a better way to predict stock moves compared to older methods and,

20
00:00:42.018 --> 00:00:42.198
well,

21
00:00:42.318 --> 00:00:43.439
other ML techniques too.

22
00:00:43.659 --> 00:00:43.779
Yeah,

23
00:00:43.780 --> 00:00:49.423
we're going to really focus on what this means for building actual trading rules and strategies,

24
00:00:49.784 --> 00:00:50.384
and importantly,

25
00:00:50.684 --> 00:00:52.606
the backtest results the authors found.

26
00:00:52.666 --> 00:00:52.826
Okay,

27
00:00:52.827 --> 00:00:53.727
so first things first,

28
00:00:53.787 --> 00:00:54.828
what data were they using?

29
00:00:55.168 --> 00:00:56.549
They used a pretty solid data set.

30
00:00:56.550 --> 00:00:59.131
It was constituents of the MSCI Japan Index.

31
00:00:59.540 --> 00:01:01.442
And they looked at 25 different factors.

32
00:01:01.862 --> 00:01:02.102
You know,

33
00:01:02.103 --> 00:01:03.143
the usual suspects,

34
00:01:03.343 --> 00:01:04.324
book-to-market ratio,

35
00:01:04.944 --> 00:01:05.645
earnings-to-price,

36
00:01:05.725 --> 00:01:06.666
past returns,

37
00:01:07.847 --> 00:01:09.008
like the last month's return.

38
00:01:09.108 --> 00:01:10.429
Just standard factors mostly.

39
00:01:10.430 --> 00:01:10.869
Pretty standard,

40
00:01:10.889 --> 00:01:11.009
yeah.

41
00:01:11.029 --> 00:01:11.389
Yeah.

42
00:01:11.429 --> 00:01:12.310
Listed in their table one.

43
00:01:12.770 --> 00:01:14.932
And the data covered a decent timeframe,

44
00:01:15.373 --> 00:01:18.075
December 1990 through November 2016.

45
00:01:18.076 --> 00:01:18.995
That's quite a stretch.

46
00:01:19.075 --> 00:01:19.195
Now,

47
00:01:19.196 --> 00:01:20.376
I saw something about a lag,

48
00:01:20.617 --> 00:01:21.757
a four-month lag.

49
00:01:21.818 --> 00:01:21.998
Yeah.

50
00:01:22.478 --> 00:01:22.678
Yes.

51
00:01:23.339 --> 00:01:24.079
That's quite important.

52
00:01:24.360 --> 00:01:25.861
They built in a four-month delay.

53
00:01:25.961 --> 00:01:26.481
Why is that?

54
00:01:27.062 --> 00:01:27.182
Well,

55
00:01:27.202 --> 00:01:28.042
it's about realism.

56
00:01:28.632 --> 00:01:30.472
The idea is even if data is released,

57
00:01:30.532 --> 00:01:32.992
it takes time for investors to actually get it,

58
00:01:33.392 --> 00:01:34.072
digest it,

59
00:01:34.172 --> 00:01:34.972
and act on it.

60
00:01:35.452 --> 00:01:38.012
So this lag tries to mimic that real-world delay.

61
00:01:38.072 --> 00:01:38.572
Makes sense.

62
00:01:38.672 --> 00:01:41.712
So the core problem was predicting next month's return.

63
00:01:42.172 --> 00:01:43.572
Using those 25 factors,

64
00:01:43.612 --> 00:01:45.032
but not just the latest values,

65
00:01:45.432 --> 00:01:47.652
they looked at the factors from the last five time points.

66
00:01:47.732 --> 00:01:48.512
Five points.

67
00:01:48.612 --> 00:01:49.312
How far apart?

68
00:01:49.552 --> 00:01:50.332
Three-month intervals.

69
00:01:50.772 --> 00:01:51.832
So looking back over what,

70
00:01:51.872 --> 00:01:53.432
about 15 months of factor history?

71
00:01:53.492 --> 00:01:53.632
Okay,

72
00:01:53.633 --> 00:01:55.692
so capturing some trend information potentially,

73
00:01:56.252 --> 00:01:57.732
and they pre-processed the data.

74
00:01:58.020 --> 00:01:58.340
They did.

75
00:01:58.400 --> 00:01:59.801
They rescaled both the inputs,

76
00:01:59.821 --> 00:02:00.442
the factors,

77
00:02:00.902 --> 00:02:01.643
and the outputs,

78
00:02:01.663 --> 00:02:02.363
the returns,

79
00:02:02.764 --> 00:02:03.364
by ranking them.

80
00:02:03.825 --> 00:02:03.965
Yeah,

81
00:02:04.045 --> 00:02:06.306
ranking across all the stocks at each point in time.

82
00:02:06.827 --> 00:02:08.068
Kind of normalizes things,

83
00:02:08.488 --> 00:02:11.911
helps the models focus on relative performance rather than absolute values.

84
00:02:12.071 --> 00:02:12.371
Got it.

85
00:02:12.691 --> 00:02:12.852
Okay,

86
00:02:12.892 --> 00:02:13.712
let's get to the good stuff.

87
00:02:13.892 --> 00:02:15.674
The trading rules and the back tests.

88
00:02:15.814 --> 00:02:17.775
They looked at different neural network setups.

89
00:02:17.976 --> 00:02:18.316
They did.

90
00:02:18.336 --> 00:02:19.977
They compared deep neural networks,

91
00:02:20.297 --> 00:02:20.938
DNNs,

92
00:02:21.518 --> 00:02:22.379
with eight layers,

93
00:02:22.619 --> 00:02:23.000
DNN8.

94
00:02:23.352 --> 00:02:24.593
and five layers DNN5.

95
00:02:25.093 --> 00:02:26.114
Against shallower ones.

96
00:02:26.115 --> 00:02:26.214
Yeah,

97
00:02:26.234 --> 00:02:26.695
exactly.

98
00:02:26.835 --> 00:02:27.916
Against a three-layer network,

99
00:02:28.156 --> 00:02:28.336
NN3.

100
00:02:29.297 --> 00:02:30.317
And they tested that

101
00:02:30.778 --> 00:02:33.660
NN3 both with and without something called dropout.

102
00:02:33.820 --> 00:02:33.980
Ah,

103
00:02:34.100 --> 00:02:34.721
dropout.

104
00:02:34.821 --> 00:02:36.222
That's for preventing overfitting,

105
00:02:36.342 --> 00:02:36.622
right?

106
00:02:36.702 --> 00:02:37.243
Precisely.

107
00:02:37.303 --> 00:02:39.685
It randomly ignores some connections during training,

108
00:02:40.125 --> 00:02:41.546
makes the network a bit more robust,

109
00:02:41.846 --> 00:02:44.028
hopefully less likely to just memorize the training data.

110
00:02:44.348 --> 00:02:47.911
So what did they find when they compared these deep versus shallow networks?

111
00:02:47.912 --> 00:02:48.632
How did they perform?

112
00:02:49.132 --> 00:02:49.292
Well,

113
00:02:49.593 --> 00:02:51.094
looking at their tables four and five,

114
00:02:51.274 --> 00:02:52.435
the general trend was

115
00:02:53.376 --> 00:02:54.337
Deeper seemed better.

116
00:02:54.397 --> 00:02:54.917
More layers,

117
00:02:54.977 --> 00:02:55.758
better predictions.

118
00:02:55.818 --> 00:02:56.138
Generally,

119
00:02:56.198 --> 00:02:56.458
yes.

120
00:02:57.059 --> 00:02:59.181
Performance measured by rank correlation,

121
00:02:59.182 --> 00:03:01.162
they called it ERR and directional accuracy,

122
00:03:01.623 --> 00:03:03.784
tended to improve as the networks got deeper.

123
00:03:04.004 --> 00:03:05.526
And COERA is?

124
00:03:06.006 --> 00:03:06.987
Rank correlation.

125
00:03:07.487 --> 00:03:07.907
Basically,

126
00:03:08.048 --> 00:03:13.272
how well the model's predicted ranking of stocks matches the actual ranking based on future returns.

127
00:03:13.712 --> 00:03:14.453
Higher is better.

128
00:03:15.013 --> 00:03:16.354
And directional accuracy?

129
00:03:16.694 --> 00:03:16.975
Simply,

130
00:03:17.435 --> 00:03:23.059
the percentage of times the model correctly guessed if a stock would go up or down relative to the median.

131
00:03:23.236 --> 00:03:23.476
Okay.

132
00:03:24.237 --> 00:03:24.977
And interestingly,

133
00:03:25.418 --> 00:03:30.161
even shallow networks with a similar number of parameters didn't quite match up to the DNNs.

134
00:03:30.302 --> 00:03:32.984
So it wasn't just about the number of knobs to turn,

135
00:03:33.024 --> 00:03:33.864
but the structure?

136
00:03:34.145 --> 00:03:34.805
It seems that way.

137
00:03:35.125 --> 00:03:36.807
The depth itself appeared to add value.

138
00:03:37.107 --> 00:03:37.287
Right.

139
00:03:37.467 --> 00:03:41.190
And the directional accuracy was statistically significant for all the network types,

140
00:03:41.430 --> 00:03:42.251
which is encouraging.

141
00:03:42.371 --> 00:03:44.092
Suggests they were definitely picking up something.

142
00:03:44.152 --> 00:03:44.353
Right.

143
00:03:44.653 --> 00:03:45.814
And the mean squared error,

144
00:03:45.874 --> 00:03:46.654
the loss function,

145
00:03:47.055 --> 00:03:49.196
was generally lower for the deeper networks too.

146
00:03:49.477 --> 00:03:49.717
Which...

147
00:03:49.888 --> 00:03:51.509
Specific models did best.

148
00:03:51.589 --> 00:03:54.632
I see DNN83 mentioned for CR.

149
00:03:54.852 --> 00:03:54.972
Yeah,

150
00:03:55.072 --> 00:03:58.635
DNN83 hit the highest CR at 0.0591.

151
00:03:58.995 --> 00:03:59.215
Now,

152
00:03:59.216 --> 00:04:01.097
0.0591,

153
00:04:01.098 --> 00:04:02.558
that sounds quite small.

154
00:04:02.658 --> 00:04:03.058
It does,

155
00:04:03.158 --> 00:04:04.079
but in this field,

156
00:04:04.339 --> 00:04:05.880
forecasting cross-sectional returns,

157
00:04:06.001 --> 00:04:08.342
even a small consistent edge like that can be meaningful.

158
00:04:08.623 --> 00:04:09.703
It's notoriously difficult.

159
00:04:09.723 --> 00:04:09.924
Okay,

160
00:04:09.964 --> 00:04:10.284
fair enough.

161
00:04:10.364 --> 00:04:12.185
And the directional accuracy for that model?

162
00:04:12.626 --> 00:04:13.126
For DNN83,

163
00:04:13.647 --> 00:04:14.948
it was about 52.6%

164
00:04:15.368 --> 00:04:16.028
for tertiles,

165
00:04:16.409 --> 00:04:17.790
splitting stocks into three groups.

166
00:04:18.216 --> 00:04:18.737
And a bit better,

167
00:04:18.837 --> 00:04:19.817
53.4%

168
00:04:20.078 --> 00:04:21.038
roughly for quintiles,

169
00:04:21.058 --> 00:04:21.919
so five groups.

170
00:04:22.099 --> 00:04:23.941
Still only slightly better than a coin flip,

171
00:04:23.942 --> 00:04:25.322
but statistically significant,

172
00:04:25.342 --> 00:04:25.642
you said.

173
00:04:25.742 --> 00:04:26.162
Exactly.

174
00:04:26.443 --> 00:04:27.143
Better than chance,

175
00:04:27.223 --> 00:04:27.784
consistently.

176
00:04:27.924 --> 00:04:28.104
Okay,

177
00:04:28.124 --> 00:04:29.525
so DNNs show promise,

178
00:04:29.526 --> 00:04:31.326
but they compared them to other ML methods too.

179
00:04:31.426 --> 00:04:31.747
They did,

180
00:04:31.767 --> 00:04:31.927
yeah.

181
00:04:32.387 --> 00:04:33.428
Crucial comparison point.

182
00:04:33.568 --> 00:04:35.069
They used support vector regression,

183
00:04:35.349 --> 00:04:35.810
SVR,

184
00:04:36.090 --> 00:04:38.552
and random forests RF as benchmarks.

185
00:04:39.152 --> 00:04:41.114
Pretty standard powerful ML techniques.

186
00:04:41.294 --> 00:04:42.955
And did they tune those models properly?

187
00:04:43.336 --> 00:04:43.736
Seems so.

188
00:04:44.096 --> 00:04:50.816
They mentioned testing 24 different setups for SVR and 37 for random forests to try and find the best configuration for each.

189
00:04:51.076 --> 00:04:53.556
So what were the best results for SVR and RF?

190
00:04:53.716 --> 00:04:54.256
For SVR,

191
00:04:54.257 --> 00:04:58.056
the best TOR was around 0.0569 directional accuracy,

192
00:04:58.216 --> 00:04:59.076
similar to the DNN,

193
00:04:59.136 --> 00:04:59.956
so maybe slightly lower,

194
00:05:00.236 --> 00:05:01.776
around 52.5%

195
00:05:01.816 --> 00:05:03.396
and 53.3%.

196
00:05:03.397 --> 00:05:04.316
And random forests?

197
00:05:04.536 --> 00:05:05.656
RF did a bit better on TOR,

198
00:05:05.657 --> 00:05:07.776
0.0576.

199
00:05:08.336 --> 00:05:10.536
Directional accuracy was quite close to the best DNNs,

200
00:05:10.576 --> 00:05:10.856
actually,

201
00:05:10.876 --> 00:05:12.516
around 52.6%

202
00:05:12.517 --> 00:05:13.816
and 53.4%.

203
00:05:13.964 --> 00:05:14.144
Okay,

204
00:05:14.205 --> 00:05:15.687
so RF looks pretty competitive there.

205
00:05:15.907 --> 00:05:18.011
How did the DNN stack up overall against these?

206
00:05:18.351 --> 00:05:19.252
This is where it gets interesting.

207
00:05:19.612 --> 00:05:21.634
Several of the top DNN patterns,

208
00:05:22.114 --> 00:05:22.835
like DNN 81,

209
00:05:22.915 --> 00:05:23.255
83,

210
00:05:23.355 --> 00:05:24.276
84,

211
00:05:24.296 --> 00:05:24.996
and 51,

212
00:05:25.477 --> 00:05:28.259
they generally outperformed SVR on most metrics.

213
00:05:28.899 --> 00:05:30.200
And against random forests,

214
00:05:30.320 --> 00:05:30.941
it was closer.

215
00:05:31.181 --> 00:05:33.143
But some DNNs showed slight advantages,

216
00:05:33.363 --> 00:05:34.544
especially in TRR.

217
00:05:35.404 --> 00:05:36.025
Specifically,

218
00:05:36.225 --> 00:05:41.149
DNN84 actually beat the best RF model on all the metrics in their comparison table,

219
00:05:41.469 --> 00:05:41.889
table six.

220
00:05:42.090 --> 00:05:42.430
Really?

221
00:05:42.530 --> 00:05:47.254
So a specific deep learning setup actually pulled ahead of a well-tuned random forest.

222
00:05:47.495 --> 00:05:48.116
In their tests,

223
00:05:48.176 --> 00:05:48.396
yes.

224
00:05:48.916 --> 00:05:52.239
It suggests maybe the deep networks were better at capturing some complex,

225
00:05:52.339 --> 00:05:55.261
nonlinear stuff in the data that even RF couldn't quite get.

226
00:05:55.541 --> 00:05:56.082
Fascinating.

227
00:05:56.522 --> 00:05:56.642
OK,

228
00:05:56.762 --> 00:05:58.063
prediction accuracy is one thing,

229
00:05:58.083 --> 00:05:59.424
but what about actual trading?

230
00:05:59.524 --> 00:06:00.986
Did they simulate a strategy?

231
00:06:01.366 --> 00:06:01.766
They did.

232
00:06:01.966 --> 00:06:03.508
A classic long-short portfolio.

233
00:06:03.568 --> 00:06:04.168
How did that work?

234
00:06:04.368 --> 00:06:04.869
Simple idea.

235
00:06:05.509 --> 00:06:07.351
Use the model's predictions to rank the stocks.

236
00:06:07.891 --> 00:06:08.832
Buy the top-ranked ones,

237
00:06:08.892 --> 00:06:09.833
sell the bottom-ranked ones.

238
00:06:10.153 --> 00:06:10.593
Equal weighting.

239
00:06:10.893 --> 00:06:11.013
OK.

240
00:06:11.514 --> 00:06:14.436
And they looked at turtle and quintile portfolios again?

241
00:06:14.607 --> 00:06:14.967
That's right.

242
00:06:14.987 --> 00:06:17.950
Splitting the stocks into three groups or five groups based on the predictions.

243
00:06:18.370 --> 00:06:19.351
One important caveat,

244
00:06:19.391 --> 00:06:19.551
though.

245
00:06:19.651 --> 00:06:20.151
Let me guess.

246
00:06:20.472 --> 00:06:21.693
Transaction costs.

247
00:06:22.013 --> 00:06:22.473
You got it.

248
00:06:22.934 --> 00:06:25.536
They didn't include transaction costs in this simulation,

249
00:06:25.996 --> 00:06:28.278
so real-world results would likely be lower.

250
00:06:28.518 --> 00:06:29.078
Understood.

251
00:06:29.839 --> 00:06:30.419
But still,

252
00:06:30.700 --> 00:06:32.441
what did the simulated performance show?

253
00:06:32.681 --> 00:06:33.802
The return versus risk?

254
00:06:34.022 --> 00:06:34.943
That's the key metric,

255
00:06:35.023 --> 00:06:35.223
right?

256
00:06:35.224 --> 00:06:36.304
The return-risk ratio.

257
00:06:36.944 --> 00:06:37.705
Looking at Table 8,

258
00:06:37.725 --> 00:06:39.126
the highest RR ratios,

259
00:06:39.426 --> 00:06:41.128
both for Turtle and Quintile portfolios,

260
00:06:41.468 --> 00:06:42.729
came from the DNN models.

261
00:06:42.989 --> 00:06:43.149
Ah,

262
00:06:43.450 --> 00:06:43.670
okay.

263
00:06:43.850 --> 00:06:44.210
Which ones?

264
00:06:44.547 --> 00:06:48.727
DNN83 had the best RR for the turtle portfolio at 1.24,

265
00:06:49.527 --> 00:06:53.287
and DNNN51 topped the quintile portfolio with 1.29.

266
00:06:53.427 --> 00:06:57.867
So the deep learning models deliver the best risk-adjusted returns in simulation.

267
00:06:58.087 --> 00:06:59.367
That's what their results suggest.

268
00:06:59.807 --> 00:07:00.547
And what's interesting is,

269
00:07:00.548 --> 00:07:00.787
remember,

270
00:07:00.867 --> 00:07:03.887
random forests sometimes had slightly higher directional accuracy.

271
00:07:04.147 --> 00:07:04.307
Yeah.

272
00:07:04.407 --> 00:07:04.527
Well,

273
00:07:04.587 --> 00:07:05.187
despite that,

274
00:07:05.547 --> 00:07:07.807
some DNNs had better risk-adjusted returns.

275
00:07:08.187 --> 00:07:12.527
It implies maybe the DNNs were better at predicting the magnitude of the returns for the winners and losers,

276
00:07:12.887 --> 00:07:13.767
not just the direction.

277
00:07:13.979 --> 00:07:16.081
That makes a difference for portfolio performance.

278
00:07:16.161 --> 00:07:16.641
Absolutely.

279
00:07:16.901 --> 00:07:19.303
Did they try combining models and an ensemble?

280
00:07:19.423 --> 00:07:19.603
Yes,

281
00:07:19.683 --> 00:07:20.064
briefly.

282
00:07:20.464 --> 00:07:22.005
They combined the best SVR,

283
00:07:22.346 --> 00:07:22.606
RF,

284
00:07:22.766 --> 00:07:23.787
and their top DNN,

285
00:07:23.927 --> 00:07:24.347
DNN83,

286
00:07:25.548 --> 00:07:26.329
with equal weights.

287
00:07:26.449 --> 00:07:27.950
Ensembles often boost performance,

288
00:07:27.990 --> 00:07:28.170
right?

289
00:07:28.630 --> 00:07:28.891
Often,

290
00:07:28.951 --> 00:07:29.111
yeah.

291
00:07:29.591 --> 00:07:31.773
The idea is different models capture different things.

292
00:07:31.893 --> 00:07:32.453
In this case,

293
00:07:32.454 --> 00:07:34.155
the ensemble did get the highest

294
00:07:34.735 --> 00:07:39.559
COR.0604, so slightly better rank prediction.

295
00:07:39.699 --> 00:07:41.260
But what about the trading simulation,

296
00:07:41.661 --> 00:07:42.241
the RR?

297
00:07:42.601 --> 00:07:43.542
That's the interesting part.

298
00:07:43.895 --> 00:07:45.076
According to Table 8,

299
00:07:45.496 --> 00:07:48.138
the improvement in directional accuracy and,

300
00:07:48.479 --> 00:07:48.939
crucially,

301
00:07:48.979 --> 00:07:53.443
the return risk ratio was pretty limited compared to the best individual DNNs.

302
00:07:53.743 --> 00:07:58.567
So the ensemble didn't quite deliver the knockout punch in terms of risk-adjusted returns here.

303
00:07:58.787 --> 00:07:59.988
Not in this specific setup,

304
00:07:59.989 --> 00:08:00.108
no.

305
00:08:00.428 --> 00:08:04.111
The best individual DNNs still held the top spots for RR.

306
00:08:04.631 --> 00:08:04.851
Okay,

307
00:08:04.891 --> 00:08:08.534
so let's try to summarize the takeaways on the trading rules and backtests.

308
00:08:08.899 --> 00:08:12.319
It seems like deeper networks generally did show an advantage here.

309
00:08:12.359 --> 00:08:12.479
Yeah,

310
00:08:12.499 --> 00:08:13.819
that seems to be the main thrust.

311
00:08:14.099 --> 00:08:19.259
Deeper networks generally perform better than shallower ones for predicting Japanese stock returns in this cross-sectional study.

312
00:08:19.619 --> 00:08:26.179
And they held their own or even slightly outperformed standard ML techniques like SVR and random forests.

313
00:08:26.339 --> 00:08:26.579
Right.

314
00:08:26.699 --> 00:08:29.699
Particularly when looking at the simulated long-short trading performance.

315
00:08:29.839 --> 00:08:29.959
Yeah.

316
00:08:30.179 --> 00:08:33.439
The DNN-based strategies yielded the highest risk-adjusted returns.

317
00:08:33.759 --> 00:08:36.659
Even though the ensemble improved correlations slightly.

318
00:08:37.115 --> 00:08:40.818
It didn't beat the best individual DNNs on that key

319
00:08:41.178 --> 00:08:41.919
RR metric.

320
00:08:41.999 --> 00:08:42.419
Correct.

321
00:08:42.599 --> 00:08:47.723
It suggests that maybe just averaging those specific models wasn't the optimal way to combine them.

322
00:08:48.424 --> 00:08:53.168
Or perhaps the individual DNNs were already capturing most of the predictable signal effectively.

323
00:08:53.268 --> 00:08:53.388
Now,

324
00:08:53.389 --> 00:08:55.389
the authors did mention limitations,

325
00:08:55.469 --> 00:08:55.730
right?

326
00:08:55.830 --> 00:08:58.212
Like focusing on standard feedforward networks.

327
00:08:58.312 --> 00:08:58.732
They did.

328
00:08:59.172 --> 00:09:01.574
They pointed out that things like recurrent neural networks,

329
00:09:01.774 --> 00:09:02.215
RNNs,

330
00:09:02.695 --> 00:09:06.658
might be interesting to explore because they're specifically designed for time series data.

331
00:09:07.275 --> 00:09:08.936
Which stock factors often are,

332
00:09:09.036 --> 00:09:10.658
even if used cross-sectionally here.

333
00:09:10.818 --> 00:09:11.318
Exactly.

334
00:09:11.438 --> 00:09:15.642
And they also suggested just exploring a wider range of deep learning architectures in general.

335
00:09:16.322 --> 00:09:19.505
There's a lot more out there now than there was even in 2018.

336
00:09:19.865 --> 00:09:21.746
So there's definitely room for more research,

337
00:09:22.047 --> 00:09:28.392
maybe exploring how these models handle different market regimes or incorporating transaction costs more explicitly.

338
00:09:28.952 --> 00:09:29.452
Absolutely.

339
00:09:29.713 --> 00:09:31.854
And I think a key question going forward is always,

340
00:09:32.715 --> 00:09:36.538
how much of this improved performance is genuinely capturing economic signal?

341
00:09:37.107 --> 00:09:41.771
versus potentially overfitting complex patterns that might not persist,

342
00:09:42.051 --> 00:09:43.112
it's the perennial challenge.

343
00:09:43.412 --> 00:09:44.393
A very important point.

344
00:09:44.533 --> 00:09:49.437
It leaves us wondering how robust these deep learning advantages truly are over the very long term.

345
00:09:49.577 --> 00:09:50.418
Definitely food for thought.

346
00:09:51.478 --> 00:09:51.699
Okay,

347
00:09:51.700 --> 00:09:53.420
well that wraps up our deep dive for today.

348
00:09:53.880 --> 00:09:56.282
Thank you for tuning in to Papers with Backtest podcast.

349
00:09:56.582 --> 00:09:58.704
We hope today's episode gave you useful insights.

350
00:09:59.044 --> 00:10:00.906
Join us next time as we break down more research.

351
00:10:01.286 --> 00:10:02.747
And for more papers and backtests,

352
00:10:02.807 --> 00:10:05.850
find us at https.paperswithbacktest.com.

353
00:10:06.030 --> 00:10:06.630
Happy trading.