WEBVTT

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

2
00:00:00.420 --> 00:00:02.783
welcome back to Papers with Backtest podcast.

3
00:00:03.143 --> 00:00:06.566
Today we dive into another algo trading research paper.

4
00:00:06.727 --> 00:00:06.906
Yeah,

5
00:00:06.988 --> 00:00:10.930
and today we're looking at one called Adaptive Moving Averages Used for Market Timing.

6
00:00:11.094 --> 00:00:11.312
Right,

7
00:00:11.352 --> 00:00:16.156
this is by Dushani Isikov and Didier Marty from the University of Freiburg,

8
00:00:17.281 --> 00:00:19.016
originally penned back in 2009,

9
00:00:19.017 --> 00:00:20.656
but then revised in 2011.

10
00:00:20.812 --> 00:00:23.281
And what's really interesting here is how it kind of builds on,

11
00:00:23.547 --> 00:00:23.703
well,

12
00:00:24.344 --> 00:00:29.125
existing research about whether technical analysis actually works.

13
00:00:29.412 --> 00:00:29.652
you know,

14
00:00:29.672 --> 00:00:30.112
makes money.

15
00:00:30.192 --> 00:00:31.433
They take it quite a bit further,

16
00:00:31.493 --> 00:00:31.893
don't they?

17
00:00:32.013 --> 00:00:32.573
They really do.

18
00:00:32.692 --> 00:00:33.954
They push things in a few ways.

19
00:00:34.192 --> 00:00:34.392
Okay,

20
00:00:34.415 --> 00:00:35.313
so let's unpack this.

21
00:00:35.415 --> 00:00:35.872
First off,

22
00:00:36.032 --> 00:00:37.352
looking at moving average rules,

23
00:00:37.411 --> 00:00:40.833
but over much longer time frames than you typically see.

24
00:00:41.114 --> 00:00:41.575
Exactly.

25
00:00:41.974 --> 00:00:45.434
Most studies cap it at maybe 200 days for the long moving average.

26
00:00:45.497 --> 00:00:46.450
These guys go up to what,

27
00:00:46.559 --> 00:00:47.122
four years?

28
00:00:48.763 --> 00:00:50.044
990 days.

29
00:00:50.059 --> 00:00:51.169
That's a huge difference.

30
00:00:51.170 --> 00:00:51.763
It makes you wonder,

31
00:00:52.153 --> 00:00:54.731
are people missing something by only looking short term?

32
00:00:54.950 --> 00:00:56.544
That seems to be the question they're asking.

33
00:00:56.825 --> 00:00:57.872
Maybe market efficiency

34
00:00:58.524 --> 00:01:00.444
looks different when you zoom way out.

35
00:01:00.725 --> 00:01:02.524
And they also dig into leverage.

36
00:01:03.044 --> 00:01:04.946
So not just if the strategy works,

37
00:01:05.005 --> 00:01:06.267
but can you boost it with,

38
00:01:06.384 --> 00:01:06.587
say,

39
00:01:07.102 --> 00:01:08.446
debt or options?

40
00:01:08.524 --> 00:01:08.790
Right.

41
00:01:08.946 --> 00:01:11.610
Which adds a whole other layer of complexity and,

42
00:01:11.649 --> 00:01:11.806
well,

43
00:01:12.048 --> 00:01:12.306
risk,

44
00:01:12.407 --> 00:01:12.610
too.

45
00:01:13.306 --> 00:01:17.485
Does leverage help or just magnify potential problems?

46
00:01:17.673 --> 00:01:18.392
And the third thing,

47
00:01:18.407 --> 00:01:19.704
which I found quite novel,

48
00:01:19.892 --> 00:01:21.563
was their market timing test.

49
00:01:22.204 --> 00:01:26.313
They developed a new way to check if these strategies actually align with,

50
00:01:26.407 --> 00:01:26.657
you know,

51
00:01:26.751 --> 00:01:26.876
the

52
00:01:27.112 --> 00:01:28.313
big picture market moves,

53
00:01:28.473 --> 00:01:29.455
bull and bear phases.

54
00:01:29.654 --> 00:01:31.435
So it's not just about the final return,

55
00:01:31.615 --> 00:01:32.697
but how you get there.

56
00:01:32.978 --> 00:01:36.963
Are you actually timing the major trends or just getting lucky on volatility?

57
00:01:37.400 --> 00:01:37.525
OK,

58
00:01:37.963 --> 00:01:39.166
so for this deep dive,

59
00:01:39.244 --> 00:01:42.345
our mission really is to get a handle on these trading rules they tested.

60
00:01:42.830 --> 00:01:44.353
Look hard at the backtest results.

61
00:01:44.431 --> 00:01:47.697
This was on the S&P 500 from 1990 to 2008.

62
00:01:47.698 --> 00:01:47.806
Yeah.

63
00:01:47.853 --> 00:01:49.260
Focusing on that profitability,

64
00:01:49.416 --> 00:01:53.197
especially for the long term rules and that market timing aspect and of course,

65
00:01:53.275 --> 00:01:54.744
how leverage changes the picture.

66
00:01:54.904 --> 00:01:55.124
Right.

67
00:01:55.125 --> 00:01:56.265
So let's start with the basics,

68
00:01:56.346 --> 00:01:57.747
the moving average rules themselves.

69
00:01:57.866 --> 00:01:58.108
Okay.

70
00:01:58.247 --> 00:02:00.550
So the core idea is pretty straightforward,

71
00:02:00.589 --> 00:02:00.811
really.

72
00:02:01.292 --> 00:02:03.792
You have a short-term moving average and a long-term one.

73
00:02:03.870 --> 00:02:05.417
And the signal comes from the crossover.

74
00:02:05.854 --> 00:02:06.253
Exactly.

75
00:02:06.472 --> 00:02:08.659
Short MA crosses above the long MA.

76
00:02:08.917 --> 00:02:10.300
That's typically your buy signal.

77
00:02:10.893 --> 00:02:11.643
Crosses below,

78
00:02:11.737 --> 00:02:12.706
that's the sell signal.

79
00:02:13.018 --> 00:02:13.596
Simple enough.

80
00:02:13.721 --> 00:02:15.128
But they mentioned a bandwidth too.

81
00:02:15.206 --> 00:02:15.909
What's that about?

82
00:02:16.112 --> 00:02:16.362
Ah,

83
00:02:16.471 --> 00:02:16.659
yeah.

84
00:02:16.800 --> 00:02:17.784
That's like a buffer zone,

85
00:02:17.846 --> 00:02:18.425
maybe 1%.

86
00:02:19.393 --> 00:02:20.268
If the MA is crossed,

87
00:02:20.269 --> 00:02:21.346
but they're still really close,

88
00:02:21.831 --> 00:02:22.784
within that band,

89
00:02:22.987 --> 00:02:24.159
the rule might just ignore it.

90
00:02:24.444 --> 00:02:24.704
Ah,

91
00:02:24.924 --> 00:02:26.944
to avoid getting whipsawed by tiny,

92
00:02:27.104 --> 00:02:28.165
meaningless moves.

93
00:02:28.245 --> 00:02:28.804
Precisely.

94
00:02:29.346 --> 00:02:32.964
Avoids too much noise and unnecessary trading when the trend isn't clear.

95
00:02:33.362 --> 00:02:35.104
So they tested a lot of these simple rules.

96
00:02:35.167 --> 00:02:35.284
Oh,

97
00:02:35.370 --> 00:02:35.807
absolutely.

98
00:02:36.026 --> 00:02:36.604
They went deep.

99
00:02:36.628 --> 00:02:38.870
1876 combinations,

100
00:02:38.886 --> 00:02:39.206
I think.

101
00:02:39.690 --> 00:02:41.940
Short MAs from one day up to 100 days.

102
00:02:42.112 --> 00:02:42.831
And the long ones.

103
00:02:42.893 --> 00:02:43.831
That's the key part.

104
00:02:44.221 --> 00:02:46.909
Long MAs from five days all the way out to that

105
00:02:47.784 --> 00:02:48.190
990-day mark.

106
00:02:48.393 --> 00:02:50.284
And they tested with and without that 1%

107
00:02:50.596 --> 00:02:50.846
band.

108
00:02:50.987 --> 00:02:51.315
Wow.

109
00:02:51.552 --> 00:02:51.752
Okay,

110
00:02:51.812 --> 00:02:53.954
so what came out of testing all those simple rules?

111
00:02:53.995 --> 00:02:55.156
Did the standard ones work?

112
00:02:55.316 --> 00:02:55.495
Well,

113
00:02:55.558 --> 00:02:55.675
no,

114
00:02:55.855 --> 00:02:56.257
not really.

115
00:02:56.296 --> 00:02:58.077
The typical short-term MA strategies,

116
00:02:58.078 --> 00:02:58.960
the ones most people know,

117
00:02:59.339 --> 00:03:00.620
they actually performed poorly.

118
00:03:01.542 --> 00:03:03.527
Many had negative returns in their tests.

119
00:03:03.581 --> 00:03:05.027
So the common wisdom didn't hold up here.

120
00:03:05.324 --> 00:03:07.003
Not in their data set for that period.

121
00:03:07.675 --> 00:03:07.910
But,

122
00:03:08.425 --> 00:03:09.253
and this is the kicker,

123
00:03:09.628 --> 00:03:12.660
the rules using those really long-term MAs,

124
00:03:12.988 --> 00:03:14.300
longer than 200 days,

125
00:03:15.019 --> 00:03:20.003
they showed average returns up to twice the buy and hold return for the S&P 500.

126
00:03:20.160 --> 00:03:20.753
Twice.

127
00:03:20.988 --> 00:03:21.108
OK,

128
00:03:21.268 --> 00:03:22.128
that's significant.

129
00:03:22.149 --> 00:03:25.151
So looking really long term might actually pay off.

130
00:03:25.330 --> 00:03:27.748
It suggests there could be slower,

131
00:03:27.791 --> 00:03:31.811
more persistent trends that maybe get overlooked because everyone's focused on the day to day.

132
00:03:32.092 --> 00:03:33.608
But these are in sample results,

133
00:03:33.670 --> 00:03:33.912
right?

134
00:03:33.951 --> 00:03:35.811
There's the risk of data snooping,

135
00:03:36.194 --> 00:03:37.850
finding patterns that aren't really there.

136
00:03:38.115 --> 00:03:38.490
Absolutely.

137
00:03:38.787 --> 00:03:40.850
That's always the caveat with initial back tests.

138
00:03:41.287 --> 00:03:42.475
The best rule on that data,

139
00:03:42.990 --> 00:03:44.287
but will it work going forward?

140
00:03:44.428 --> 00:03:47.490
So did they check how consistent this was over time?

141
00:03:47.506 --> 00:03:48.459
They broke the period down,

142
00:03:48.475 --> 00:03:48.819
didn't they?

143
00:03:48.990 --> 00:03:49.287
They did.

144
00:03:49.516 --> 00:03:52.800
Five sub-periods from 1990 right through to 2008.

145
00:03:53.179 --> 00:03:54.561
And the story was pretty consistent.

146
00:03:54.581 --> 00:03:56.143
The short-term rules stayed bad.

147
00:03:56.202 --> 00:03:56.362
Yep.

148
00:03:56.683 --> 00:03:58.143
Consistently poor across the board.

149
00:03:58.643 --> 00:03:59.745
But the long-term ones,

150
00:04:00.667 --> 00:04:03.190
their performance didn't really seem to degrade over time.

151
00:04:03.511 --> 00:04:04.393
Even in the later periods,

152
00:04:04.394 --> 00:04:06.253
like the run-up to the 08 crisis?

153
00:04:06.565 --> 00:04:07.300
Especially then,

154
00:04:07.440 --> 00:04:07.737
actually.

155
00:04:08.128 --> 00:04:11.893
They generally beat buy and hold in those later sub-periods.

156
00:04:12.237 --> 00:04:12.643
And get this,

157
00:04:12.753 --> 00:04:13.565
during that last period,

158
00:04:13.643 --> 00:04:14.222
2006,

159
00:04:14.253 --> 00:04:14.847
2008,

160
00:04:15.284 --> 00:04:16.722
which included the Lehman collapse.

161
00:04:16.925 --> 00:04:17.128
Yeah.

162
00:04:17.580 --> 00:04:19.021
That was brutal for buy and hold.

163
00:04:19.162 --> 00:04:19.441
Right.

164
00:04:19.643 --> 00:04:20.684
Buy and hold was negative.

165
00:04:21.203 --> 00:04:22.523
But most of the EMA rules,

166
00:04:22.703 --> 00:04:23.828
except the very short ones,

167
00:04:24.266 --> 00:04:25.969
actually generated positive returns.

168
00:04:26.109 --> 00:04:26.547
Wow.

169
00:04:26.812 --> 00:04:30.195
So those longer trends maybe offered some protection in the crash.

170
00:04:30.453 --> 00:04:31.195
It seems that way.

171
00:04:31.656 --> 00:04:33.156
Suggests a certain robustness.

172
00:04:33.891 --> 00:04:34.016
So,

173
00:04:34.109 --> 00:04:34.344
okay,

174
00:04:34.797 --> 00:04:36.641
these simple long-term rules show promise,

175
00:04:37.016 --> 00:04:39.438
but they're still prone to that data snooping issue.

176
00:04:39.844 --> 00:04:41.891
How did they try to build something more robust?

177
00:04:42.297 --> 00:04:42.938
Good question.

178
00:04:43.312 --> 00:04:46.328
They moved on to four more complex adaptive strategies.

179
00:04:46.552 --> 00:04:47.372
trying to overcome,

180
00:04:47.613 --> 00:04:47.872
you know,

181
00:04:48.113 --> 00:04:49.852
just picking one lucky rule.

182
00:04:49.992 --> 00:04:50.194
Okay,

183
00:04:50.213 --> 00:04:50.614
what were they?

184
00:04:50.852 --> 00:04:52.332
First was OPTAL.

185
00:04:52.895 --> 00:04:56.817
This one basically looked back every single day at the entire history up to that point.

186
00:04:56.957 --> 00:04:59.434
And picked the best simple MA rule from the past.

187
00:04:59.637 --> 00:05:00.114
Exactly.

188
00:05:00.137 --> 00:05:03.371
It picked the historical best performer and used its signal for that day.

189
00:05:03.715 --> 00:05:05.731
Started doing this from 1994 onwards,

190
00:05:06.043 --> 00:05:07.012
constantly adapting.

191
00:05:07.121 --> 00:05:07.278
Okay,

192
00:05:07.340 --> 00:05:08.403
constantly re-optimizing.

193
00:05:08.450 --> 00:05:08.778
What else?

194
00:05:09.135 --> 00:05:09.796
Then there was OPT4,

195
00:05:10.416 --> 00:05:11.116
a bit different.

196
00:05:11.697 --> 00:05:15.181
It picked the best simple rule over a four-year selection period.

197
00:05:15.579 --> 00:05:17.540
And then used that rule for the next four years.

198
00:05:17.603 --> 00:05:17.861
Right.

199
00:05:18.064 --> 00:05:19.845
Applied it for a four-year evaluation period.

200
00:05:20.267 --> 00:05:21.486
And they repeated this cycle,

201
00:05:21.564 --> 00:05:22.345
select for four,

202
00:05:22.704 --> 00:05:23.626
evaluate for four,

203
00:05:23.751 --> 00:05:24.493
across their data.

204
00:05:24.587 --> 00:05:26.408
Kind of a rolling out of sample approach.

205
00:05:26.470 --> 00:05:26.798
Got it.

206
00:05:27.220 --> 00:05:28.314
Less frequent adaptation.

207
00:05:28.315 --> 00:05:28.908
What were the other two?

208
00:05:29.251 --> 00:05:32.673
They involved more of a consensus approach.

209
00:05:33.017 --> 00:05:33.876
Vote was one.

210
00:05:34.173 --> 00:05:38.361
It looked at all the simple rules that beat the market in the previous selection period.

211
00:05:39.003 --> 00:05:43.610
And then the trading signal for the day was whatever the majority of those winning rules were saying,

212
00:05:43.949 --> 00:05:44.289
buy,

213
00:05:44.649 --> 00:05:44.910
sell,

214
00:05:45.008 --> 00:05:45.531
or neutral.

215
00:05:45.852 --> 00:05:47.071
Like wisdom of the crowd,

216
00:05:47.453 --> 00:05:49.656
but only the historically successful crowd.

217
00:05:49.750 --> 00:05:50.235
Interesting.

218
00:05:50.594 --> 00:05:52.360
A democratic approach among strategies.

219
00:05:52.875 --> 00:05:53.461
And the last one.

220
00:05:53.860 --> 00:05:54.336
Partial.

221
00:05:55.039 --> 00:05:56.055
Similar idea to vote.

222
00:05:56.321 --> 00:05:58.008
Finding the rules that worked in the prior period.

223
00:05:58.102 --> 00:05:59.899
But instead of just taking the majority signal,

224
00:06:00.258 --> 00:06:01.961
it calculated a fractional position.

225
00:06:02.024 --> 00:06:02.508
How'd it work?

226
00:06:02.789 --> 00:06:03.946
It averaged the signals,

227
00:06:04.055 --> 00:06:05.211
treating sell as minus one,

228
00:06:05.446 --> 00:06:06.289
neutral as zero,

229
00:06:06.789 --> 00:06:07.508
buy as one.

230
00:06:07.930 --> 00:06:08.336
So if...

231
00:06:08.783 --> 00:06:09.043
say

232
00:06:09.403 --> 00:06:11.604
70% of rules said by and 30%

233
00:06:11.605 --> 00:06:12.104
said neutral,

234
00:06:12.405 --> 00:06:14.545
you might be like 0.7 long.

235
00:06:14.983 --> 00:06:15.163
Ah,

236
00:06:15.405 --> 00:06:18.268
so it scales the position based on the strength of the consensus.

237
00:06:18.448 --> 00:06:19.088
They're nuanced.

238
00:06:19.307 --> 00:06:19.885
Exactly.

239
00:06:20.424 --> 00:06:20.627
Okay,

240
00:06:20.666 --> 00:06:22.565
so these are the four complex strategies.

241
00:06:23.049 --> 00:06:27.049
How did they actually stack up when back-tested from 1994 to 2008?

242
00:06:27.549 --> 00:06:29.143
They compared them against buy and hold,

243
00:06:29.424 --> 00:06:29.831
BH,

244
00:06:30.268 --> 00:06:31.206
a random walk,

245
00:06:31.393 --> 00:06:31.846
or W,

246
00:06:32.409 --> 00:06:34.049
and the single best in-sample,

247
00:06:34.362 --> 00:06:34.721
simple.

248
00:06:35.487 --> 00:06:35.721
Right,

249
00:06:35.846 --> 00:06:36.674
the moment of truth.

250
00:06:37.346 --> 00:06:37.846
What about just

251
00:06:38.191 --> 00:06:39.753
directional accuracy first,

252
00:06:40.073 --> 00:06:41.274
getting buys and sells right.

253
00:06:41.475 --> 00:06:41.594
Well,

254
00:06:41.595 --> 00:06:42.495
for buy signals,

255
00:06:42.534 --> 00:06:45.940
they were slightly better than just being long all the time via buy and hold.

256
00:06:46.323 --> 00:06:48.604
But where they really made a difference was on the sell side.

257
00:06:48.698 --> 00:06:50.581
Better than just shorting the market randomly.

258
00:06:50.706 --> 00:06:51.604
Significantly better.

259
00:06:51.667 --> 00:06:51.823
Yeah.

260
00:06:52.221 --> 00:06:57.604
Which suggests they had some real ability to identify downturns or at least times not to be long.

261
00:06:57.948 --> 00:06:58.073
OK,

262
00:06:58.448 --> 00:06:58.917
makes sense.

263
00:06:59.510 --> 00:07:00.495
But what about the bottom line,

264
00:07:00.760 --> 00:07:01.401
the returns?

265
00:07:01.557 --> 00:07:02.760
That's where it gets really compelling.

266
00:07:03.120 --> 00:07:07.557
Buy and hold returned about 6.16 percent annualized noon return over that period.

267
00:07:07.647 --> 00:07:08.949
And the complex strategies.

268
00:07:09.249 --> 00:07:10.509
All four beat it soundly.

269
00:07:10.890 --> 00:07:12.673
The lowest was OPT all at around

270
00:07:13.111 --> 00:07:18.138
10.7%, and the highest was OPT4 at a pretty impressive 14.61%.

271
00:07:18.139 --> 00:07:18.255
Wow,

272
00:07:18.357 --> 00:07:20.998
OPT4 even beat the single best in sample rule.

273
00:07:21.341 --> 00:07:21.662
Slightly,

274
00:07:21.740 --> 00:07:21.857
yeah.

275
00:07:22.201 --> 00:07:24.607
OPT4 hit 14.61%,

276
00:07:24.608 --> 00:07:26.482
best was 14.33%.

277
00:07:27.248 --> 00:07:28.029
And the random walk,

278
00:07:28.060 --> 00:07:28.232
well,

279
00:07:28.295 --> 00:07:28.826
it lost money,

280
00:07:28.888 --> 00:07:30.576
about negative 8.6%

281
00:07:30.577 --> 00:07:31.107
annualized.

282
00:07:31.545 --> 00:07:34.966
So these complex adaptive rules really delivered on the return front.

283
00:07:35.083 --> 00:07:36.444
What about compounded returns?

284
00:07:36.504 --> 00:07:38.125
That's often more important for investors.

285
00:07:38.227 --> 00:07:39.149
Even better story there,

286
00:07:39.348 --> 00:07:40.149
relatively speaking.

287
00:07:40.547 --> 00:07:43.532
Buy and hold compounded at only 4.41%

288
00:07:43.594 --> 00:07:43.969
annually.

289
00:07:44.274 --> 00:07:45.290
The complex strategies.

290
00:07:45.415 --> 00:07:47.032
They range from 9.25%

291
00:07:47.079 --> 00:07:50.258
for OPT all up to 13.62%

292
00:07:50.259 --> 00:07:51.040
for OPT for.

293
00:07:51.157 --> 00:07:53.938
That's a huge difference over 15 years.

294
00:07:54.094 --> 00:07:54.563
Absolutely.

295
00:07:54.594 --> 00:07:55.844
And here's something else striking.

296
00:07:56.391 --> 00:07:57.391
They didn't trade much.

297
00:07:57.688 --> 00:07:57.985
Really?

298
00:07:58.204 --> 00:07:59.266
With all that adapting,

299
00:07:59.313 --> 00:08:00.625
especially OPT all?

300
00:08:00.750 --> 00:08:01.329
Surprisingly,

301
00:08:01.391 --> 00:08:01.516
no.

302
00:08:02.011 --> 00:08:03.272
Over the whole 15 years,

303
00:08:03.313 --> 00:08:06.797
the complex strategies only made between 7 and 39 trades in total.

304
00:08:07.055 --> 00:08:07.817
7 trades.

305
00:08:08.137 --> 00:08:08.996
Over 15 years.

306
00:08:08.997 --> 00:08:10.141
That's incredibly infrequent.

307
00:08:10.219 --> 00:08:10.704
It really is.

308
00:08:10.977 --> 00:08:12.461
Compare that to the random walk strategy,

309
00:08:12.539 --> 00:08:13.704
nearly 4,000 trades.

310
00:08:13.844 --> 00:08:18.000
So transaction costs wouldn't be a major issue for these complex MA strategies.

311
00:08:18.001 --> 00:08:18.750
Very likely not.

312
00:08:19.235 --> 00:08:21.391
The paper calculated the break-even transaction costs,

313
00:08:21.407 --> 00:08:24.829
how high costs would need to be to wipe out the excess profits over buy and hold.

314
00:08:25.094 --> 00:08:25.797
They were quite high,

315
00:08:26.235 --> 00:08:28.297
ranging from 1.74%

316
00:08:28.298 --> 00:08:29.000
up to over 12%

317
00:08:29.375 --> 00:08:31.000
per round trip for some strategies.

318
00:08:31.411 --> 00:08:32.773
way higher than typical costs,

319
00:08:33.212 --> 00:08:35.134
suggests the profitability is pretty robust.

320
00:08:35.195 --> 00:08:35.357
Okay,

321
00:08:35.435 --> 00:08:36.236
high returns,

322
00:08:36.556 --> 00:08:37.576
low trading frequency,

323
00:08:37.798 --> 00:08:38.759
robust to cost.

324
00:08:39.177 --> 00:08:40.802
But did they just take on massive risk?

325
00:08:40.880 --> 00:08:42.279
What about risk-adjusted returns?

326
00:08:42.365 --> 00:08:42.880
Good point.

327
00:08:43.060 --> 00:08:45.482
They looked at Jensen's alpha and Sharpe ratios.

328
00:08:46.123 --> 00:08:50.193
The alphas were positive and generally larger than buy-in holds,

329
00:08:50.755 --> 00:08:52.505
some statistically significant too.

330
00:08:52.771 --> 00:08:54.677
Meaning they generated excess returns,

331
00:08:54.787 --> 00:08:56.177
even accounting for market risk.

332
00:08:56.271 --> 00:08:56.599
Correct.

333
00:08:56.880 --> 00:08:57.693
And the Sharpe ratios,

334
00:08:57.724 --> 00:08:59.677
which measure return per unit of risk.

335
00:09:00.159 --> 00:09:03.783
They were at least twice as high for the complex strategies compared to buy and hold.

336
00:09:04.042 --> 00:09:06.927
So better returns and better risk-adjusted performance.

337
00:09:06.986 --> 00:09:07.966
That's pretty compelling.

338
00:09:07.967 --> 00:09:08.466
It really is.

339
00:09:08.669 --> 00:09:12.388
It suggests the higher returns weren't just a reward for taking on tons more volatility.

340
00:09:12.615 --> 00:09:15.755
And when they looked inside the winning strategies like OPTL and OPT4,

341
00:09:16.162 --> 00:09:18.693
did they confirm the importance of those long-term MAs?

342
00:09:19.021 --> 00:09:19.505
Absolutely.

343
00:09:19.974 --> 00:09:23.037
The rules selected by OPTL often had long MAs around,

344
00:09:23.099 --> 00:09:23.537
like 6,

345
00:09:23.615 --> 00:09:24.005
15,

346
00:09:24.099 --> 00:09:24.224
6,

347
00:09:24.225 --> 00:09:25.130
165,

348
00:09:25.208 --> 00:09:26.412
even 940 days.

349
00:09:26.959 --> 00:09:28.462
OPT4 also favored long MA665-515,

350
00:09:28.463 --> 00:09:31.768
415 days in different periods.

351
00:09:32.108 --> 00:09:33.249
And the best simple rule.

352
00:09:33.549 --> 00:09:36.858
Also had a long MA465 days.

353
00:09:37.499 --> 00:09:38.483
So the theme holds.

354
00:09:39.241 --> 00:09:43.725
Looking way beyond the standard 200 days seems crucial to the success they found.

355
00:09:44.142 --> 00:09:44.342
Okay,

356
00:09:44.382 --> 00:09:45.644
so profitability looks good.

357
00:09:45.964 --> 00:09:48.285
What about that market timing aspect they wanted to test?

358
00:09:48.348 --> 00:09:51.109
Did the strategies actually align with bull and bear markets?

359
00:09:51.469 --> 00:09:51.711
Right.

360
00:09:51.730 --> 00:09:56.973
They used their new test based on identifying long-term market phases using an algorithm adapted from

361
00:09:57.355 --> 00:09:58.434
Pagan and Sosanoff.

362
00:09:58.676 --> 00:10:00.879
And did the rules stay long in bull markets?

363
00:10:00.957 --> 00:10:01.426
Pretty much.

364
00:10:01.613 --> 00:10:04.863
The complex rules were in a buy state for over 90%

365
00:10:04.864 --> 00:10:06.910
of the days identified as bull market phases.

366
00:10:07.051 --> 00:10:07.270
Okay,

367
00:10:07.301 --> 00:10:08.082
that's good alignment.

368
00:10:08.207 --> 00:10:09.051
What about bear markets?

369
00:10:09.066 --> 00:10:10.332
Did they get out or go short?

370
00:10:10.730 --> 00:10:12.792
They were in a sell state less consistently there,

371
00:10:12.891 --> 00:10:14.274
maybe around 61%

372
00:10:14.275 --> 00:10:14.934
to 74%

373
00:10:15.055 --> 00:10:16.036
of the bear market days.

374
00:10:16.317 --> 00:10:16.876
Not perfect,

375
00:10:17.137 --> 00:10:19.555
but still much better than chance or just staying long.

376
00:10:19.758 --> 00:10:20.798
So overall alignment,

377
00:10:20.923 --> 00:10:24.001
the total percentage of days the signals matched the market phase.

378
00:10:24.204 --> 00:10:25.548
That ranged from 83%

379
00:10:25.626 --> 00:10:26.469
to 88%

380
00:10:26.470 --> 00:10:27.548
for the complex rules.

381
00:10:27.923 --> 00:10:29.048
Better than buy and hold,

382
00:10:29.141 --> 00:10:30.032
which was right about

383
00:10:30.407 --> 00:10:32.251
75% of the time by definition,

384
00:10:32.673 --> 00:10:33.173
always long.

385
00:10:33.376 --> 00:10:35.594
And they checked if this was statistically significant,

386
00:10:35.688 --> 00:10:36.438
not just luck.

387
00:10:36.829 --> 00:10:37.063
Yes,

388
00:10:37.298 --> 00:10:38.673
they used bootstrap analysis.

389
00:10:39.338 --> 00:10:41.800
And it confirmed the results were statistically significant.

390
00:10:42.562 --> 00:10:45.304
It suggests a genuine ability to time these broad,

391
00:10:45.425 --> 00:10:46.644
long-term market cycles.

392
00:10:46.925 --> 00:10:47.523
Interesting.

393
00:10:47.968 --> 00:10:52.468
So profitable strategies with statistically significant market timing ability.

394
00:10:52.929 --> 00:10:53.070
Now,

395
00:10:53.288 --> 00:10:54.437
what about adding leverage?

396
00:10:54.749 --> 00:10:55.937
Did that supercharge things?

397
00:10:56.171 --> 00:10:57.343
They looked at debt first.

398
00:10:57.468 --> 00:10:57.640
Yeah,

399
00:10:57.812 --> 00:10:59.015
borrowing money to invest more.

400
00:10:59.406 --> 00:11:01.359
They simulated borrowing at different rates,

401
00:11:01.374 --> 00:11:02.265
the US prime rate,

402
00:11:02.640 --> 00:11:03.718
the risk-free rate,

403
00:11:04.062 --> 00:11:04.343
and even,

404
00:11:04.562 --> 00:11:05.187
hypothetically,

405
00:11:05.562 --> 00:11:06.281
at no cost.

406
00:11:06.742 --> 00:11:09.045
And how did that affect the complex strategies?

407
00:11:09.103 --> 00:11:10.967
It significantly boosted the returns,

408
00:11:11.389 --> 00:11:13.670
both mean simple returns and importantly,

409
00:11:13.869 --> 00:11:15.193
the compounded returns,

410
00:11:15.709 --> 00:11:17.850
even when factoring in realistic borrowing costs.

411
00:11:17.951 --> 00:11:18.990
Any standout numbers?

412
00:11:19.451 --> 00:11:19.576
Well,

413
00:11:19.615 --> 00:11:19.896
OPT4,

414
00:11:20.334 --> 00:11:21.693
which was already the best performer,

415
00:11:21.975 --> 00:11:24.818
showed really impressive compounded returns with debt leverage,

416
00:11:25.818 --> 00:11:29.006
potentially up to 24 percent annualized if you could borrow for free,

417
00:11:29.568 --> 00:11:31.771
but still substantially higher even with costs.

418
00:11:31.943 --> 00:11:33.975
How did leveraged buy and hold compare?

419
00:11:34.053 --> 00:11:34.975
Did it get the same?

420
00:11:36.107 --> 00:11:37.848
It also saw higher mean returns,

421
00:11:37.928 --> 00:11:40.350
but its compounded returns didn't improve nearly as much,

422
00:11:40.412 --> 00:11:41.834
especially with borrowing costs.

423
00:11:42.393 --> 00:11:43.455
This highlights something important.

424
00:11:43.795 --> 00:11:44.194
What's that?

425
00:11:44.537 --> 00:11:49.522
It suggests the leverage gains in the complex strategies weren't just due to the leverage itself,

426
00:11:49.944 --> 00:11:53.037
but because the underlying strategies had actual forecasting ability,

427
00:11:53.490 --> 00:11:57.022
they avoided some of the big drawdowns that hurt leveraged buy and hold.

428
00:11:57.209 --> 00:11:57.412
Right.

429
00:11:57.444 --> 00:11:58.881
Leverage magnifies losses too.

430
00:11:59.490 --> 00:12:01.490
And the risk-adjusted performance,

431
00:12:01.647 --> 00:12:02.100
alphas.

432
00:12:02.386 --> 00:12:08.951
The alphas for the leveraged complex strategies were substantially higher than both the unleveraged versions and leveraged buy and hold.

433
00:12:09.232 --> 00:12:09.353
OK,

434
00:12:09.451 --> 00:12:11.654
so debt leverage seemed to work well with these strategies.

435
00:12:11.795 --> 00:12:12.779
What about options?

436
00:12:12.818 --> 00:12:15.240
That's usually seen as a more aggressive way to leverage.

437
00:12:15.818 --> 00:12:16.076
Right.

438
00:12:16.303 --> 00:12:18.701
They simulated allocating a portion of the capital,

439
00:12:18.795 --> 00:12:19.139
5%,

440
00:12:19.357 --> 00:12:19.951
10%,

441
00:12:20.185 --> 00:12:21.123
or 15%,

442
00:12:21.498 --> 00:12:25.826
to buying exchange traded call or put options based on the strategy signal.

443
00:12:25.827 --> 00:12:26.326
And the result?

444
00:12:26.810 --> 00:12:27.795
High octane returns.

445
00:12:27.951 --> 00:12:29.154
High octane volatility,

446
00:12:29.232 --> 00:12:29.639
definitely.

447
00:12:30.062 --> 00:12:36.166
The options leverage led to huge swings and massive differences between the simple average returns and the compounded returns.

448
00:12:36.167 --> 00:12:39.729
So did it actually improve the compounded returns in the end?

449
00:12:39.791 --> 00:12:40.572
It was mixed.

450
00:12:40.994 --> 00:12:41.471
For OPT4,

451
00:12:42.072 --> 00:12:43.416
allocating up to 15%

452
00:12:43.417 --> 00:12:45.900
in options did boost the compounded return,

453
00:12:46.111 --> 00:12:47.221
reaching over 19%

454
00:12:47.361 --> 00:12:47.924
annualized.

455
00:12:48.299 --> 00:12:49.299
But for OPT all,

456
00:12:49.908 --> 00:12:51.221
the performance actually got worse,

457
00:12:51.627 --> 00:12:54.346
turning negative with the higher option allocations.

458
00:12:54.768 --> 00:12:56.690
The volatility drag was just too much.

459
00:12:57.065 --> 00:12:59.424
So the conclusion on leverage was?

460
00:12:59.838 --> 00:13:05.283
Debt leverage seemed much more suitable for enhancing these particular long-term MA strategies.

461
00:13:05.622 --> 00:13:09.966
It provided a significant boost without the extreme volatility introduced by options.

462
00:13:10.310 --> 00:13:12.708
Did they look at any other risk aspects,

463
00:13:12.794 --> 00:13:13.794
like downside risk?

464
00:13:14.115 --> 00:13:14.435
Briefly,

465
00:13:14.490 --> 00:13:14.630
yeah.

466
00:13:14.896 --> 00:13:16.794
They mentioned things like downside upside betas,

467
00:13:17.029 --> 00:13:18.060
Soutino ratios,

468
00:13:18.372 --> 00:13:19.263
even cost skewness.

469
00:13:19.763 --> 00:13:22.279
The results generally suggested that complex strategies,

470
00:13:22.466 --> 00:13:23.622
especially with debt leverage,

471
00:13:23.904 --> 00:13:25.966
didn't necessarily increase downside risk.

472
00:13:26.242 --> 00:13:27.303
portionally to the returns.

473
00:13:27.443 --> 00:13:30.045
Maybe even offered some protection against negative skewness.

474
00:13:30.385 --> 00:13:30.928
Potentially,

475
00:13:30.967 --> 00:13:31.107
yeah.

476
00:13:31.408 --> 00:13:33.631
Like skewness insurance without giving up the returns,

477
00:13:33.650 --> 00:13:34.693
which is quite desirable.

478
00:13:34.787 --> 00:13:34.912
OK,

479
00:13:35.029 --> 00:13:36.092
so let's try and sum this up.

480
00:13:36.131 --> 00:13:38.217
What's the big picture takeaway from this paper?

481
00:13:38.576 --> 00:13:38.717
Well,

482
00:13:39.131 --> 00:13:42.381
I think the main point is that these complex moving average rules,

483
00:13:42.412 --> 00:13:45.693
especially the ones using those unusually long look back periods,

484
00:13:46.162 --> 00:13:47.443
seem to show significant,

485
00:13:47.521 --> 00:13:50.162
robust profitability and market timing ability,

486
00:13:50.803 --> 00:13:55.209
at least on the S&P 500 during their test period 1994-2008.

487
00:13:55.382 --> 00:13:59.122
Which definitely challenges the simple view of the efficient market hypothesis,

488
00:13:59.182 --> 00:13:59.423
right?

489
00:13:59.462 --> 00:13:59.864
It does.

490
00:14:00.282 --> 00:14:04.063
Particularly the idea that simple technical rules can't beat the market.

491
00:14:04.907 --> 00:14:06.907
These findings suggest maybe they can,

492
00:14:07.602 --> 00:14:10.204
if you look long-term enough and combine them intelligently.

493
00:14:10.907 --> 00:14:15.251
And the fact that debt leverage amplified these profits further strengthens the case.

494
00:14:15.766 --> 00:14:20.235
Why might these long-term strategies work when shorter-term ones didn't seem to?

495
00:14:20.923 --> 00:14:21.048
The

496
00:14:21.718 --> 00:14:25.778
Paper kind of hints that maybe most market participants are just too focused on the short term,

497
00:14:26.321 --> 00:14:27.005
on the noise.

498
00:14:27.511 --> 00:14:32.178
They might be creating inefficiencies or trends over longer horizons that aren't being fully exploited.

499
00:14:32.357 --> 00:14:34.139
Herd behavior focused on the near term,

500
00:14:34.178 --> 00:14:34.639
perhaps.

501
00:14:34.896 --> 00:14:35.197
Could be.

502
00:14:35.779 --> 00:14:37.381
Or maybe institutional constraints,

503
00:14:37.521 --> 00:14:38.521
performance pressures.

504
00:14:39.045 --> 00:14:39.982
It's hard to say for sure.

505
00:14:40.146 --> 00:14:42.545
The authors themselves suggest more research is needed.

506
00:14:42.724 --> 00:14:43.068
Like what?

507
00:14:43.389 --> 00:14:49.326
Developing maybe even stronger statistical tests to really confirm these findings aren't just elaborate data mining.

508
00:14:49.717 --> 00:14:52.029
And testing these ideas over longer time periods and,

509
00:14:52.076 --> 00:14:52.498
crucially,

510
00:14:52.795 --> 00:14:53.576
in other markets too.

511
00:14:53.998 --> 00:14:56.873
Does this work for bonds or commodities or international stocks?

512
00:14:57.235 --> 00:14:57.455
Right.

513
00:14:57.815 --> 00:14:58.796
Lots more to explore.

514
00:14:59.455 --> 00:15:04.998
But definitely a thought-provoking study suggesting value and looking beyond the usual technical analysis horizons.

515
00:15:05.154 --> 00:15:05.678
Absolutely.

516
00:15:06.357 --> 00:15:07.920
A good reminder to sometimes zoom out.

517
00:15:08.396 --> 00:15:08.756
Indeed.

518
00:15:09.576 --> 00:15:09.701
Well,

519
00:15:09.702 --> 00:15:11.420
that brings us to the end of this deep dive.

520
00:15:11.521 --> 00:15:14.037
Thank you for tuning in to Papers with Backtests podcast.

521
00:15:14.084 --> 00:15:16.240
We hope today's episode gave you useful insights.

522
00:15:16.724 --> 00:15:18.646
Join us next time as we break down more research.

523
00:15:18.881 --> 00:15:20.224
And for more papers and backtests,

524
00:15:20.256 --> 00:15:23.256
find us at https.paperswithbacktests.com.

525
00:15:23.803 --> 00:15:24.365
Happy trading.