WEBVTT

1
00:00:00.305 --> 00:00:02.205
The market keeps
setting record highs.

2
00:00:02.925 --> 00:00:04.206
Should you wait for a dip

3
00:00:04.206 --> 00:00:04.975
before you buy?

4
00:00:05.842 --> 00:00:07.126
The study needs a target

5
00:00:07.126 --> 00:00:08.142
and a fixed window.

6
00:00:08.842 --> 00:00:11.062
The target is the
same broad stock fund.

7
00:00:11.582 --> 00:00:15.362
15 years, from
2010 through 2024.

8
00:00:16.082 --> 00:00:17.610
Fixing the dates
is what makes

9
00:00:17.610 --> 00:00:18.832
the result reproducible.

10
00:00:19.489 --> 00:00:22.018
Signal Analyzer compares
a predictor with the

11
00:00:22.018 --> 00:00:24.209
outcome that followed
each observation.

12
00:00:25.189 --> 00:00:26.934
It measures
relationship strength

13
00:00:26.934 --> 00:00:28.838
and checks whether
that relationship

14
00:00:28.838 --> 00:00:29.949
persists across eras.

15
00:00:30.869 --> 00:00:33.186
As an example, we will
test whether a broad

16
00:00:33.186 --> 00:00:34.964
stock fund's price
level predicts

17
00:00:34.964 --> 00:00:36.149
its next-month return.

18
00:00:37.893 --> 00:00:40.153
The predictor is the
fund's price level.

19
00:00:40.853 --> 00:00:43.350
Another study could
use a moving average,

20
00:00:43.350 --> 00:00:44.994
a trailing return, a yield,

21
00:00:44.994 --> 00:00:46.273
or a calendar effect.

22
00:00:47.093 --> 00:00:48.923
A predictor can
also combine several

23
00:00:48.923 --> 00:00:50.193
of those transformations.

24
00:00:51.839 --> 00:00:54.454
The outcome it has to
predict is the return

25
00:00:54.454 --> 00:00:56.279
over the next
21 trading days.

26
00:00:56.779 --> 00:00:57.979
Roughly one month ahead.

27
00:00:59.534 --> 00:01:01.911
Sampling and splits
decide what counts

28
00:01:01.911 --> 00:01:02.974
as evidence here.

29
00:01:03.694 --> 00:01:06.652
Spacing decides whether
two observations overlap,

30
00:01:06.652 --> 00:01:08.585
and overlapping
observations are

31
00:01:08.585 --> 00:01:10.094
not independent evidence.

32
00:01:10.874 --> 00:01:13.466
Predictor lag decides
how much of the past the

33
00:01:13.466 --> 00:01:14.874
signal is allowed to see.

34
00:01:15.594 --> 00:01:17.590
The monthly samples
are divided into

35
00:01:17.590 --> 00:01:18.754
3 historical periods.

36
00:01:20.564 --> 00:01:23.344
Fit settings control how
the relationship is drawn.

37
00:01:24.244 --> 00:01:26.114
Choose a smooth
polynomial curve,

38
00:01:26.114 --> 00:01:28.664
or bins showing the
average outcome by group.

39
00:01:29.424 --> 00:01:31.808
This study uses the
default smooth curve,

40
00:01:31.808 --> 00:01:33.377
which is worth remembering:

41
00:01:33.377 --> 00:01:35.238
a curve can look
convincing even

42
00:01:35.238 --> 00:01:36.924
when the
correlation is zero.

43
00:01:37.784 --> 00:01:40.324
Does price level predict
the next month's return?

44
00:01:41.124 --> 00:01:43.242
Run 15 years of
monthly samples with

45
00:01:43.242 --> 00:01:45.124
non-overlapping
forward windows.

46
00:01:45.938 --> 00:01:48.038
First, understand
the threshold sweep.

47
00:01:48.798 --> 00:01:50.838
Each cutoff splits
the same sample

48
00:01:50.838 --> 00:01:52.458
and compares outcomes above

49
00:01:52.458 --> 00:01:53.238
and below it.

50
00:01:53.798 --> 00:01:56.171
The same observations
are tested against

51
00:01:56.171 --> 00:01:57.298
9 threshold values.

52
00:01:58.078 --> 00:02:00.314
Repeated tests can
produce an attractive gap

53
00:02:00.314 --> 00:02:01.178
from noise alone.

54
00:02:02.158 --> 00:02:03.485
Read the observation count

55
00:02:03.485 --> 00:02:04.608
before the widest gap.

56
00:02:05.650 --> 00:02:08.250
179 independent
observations.

57
00:02:09.230 --> 00:02:11.221
The correlation
between price level

58
00:02:11.221 --> 00:02:13.325
and the next month's
return is 0.000,

59
00:02:13.325 --> 00:02:14.690
to three decimal places.

60
00:02:15.948 --> 00:02:17.728
On the chart that
is a flat line.

61
00:02:18.148 --> 00:02:20.785
Where the market sits tells
you almost nothing about

62
00:02:20.785 --> 00:02:21.748
where it goes next.

63
00:02:23.697 --> 00:02:25.647
The sweep is where
research goes wrong.

64
00:02:26.417 --> 00:02:30.012
Below the cutoff at 80%
of the price distribution,

65
00:02:30.012 --> 00:02:32.457
the forward return
averaged 1.28%.

66
00:02:32.457 --> 00:02:37.117
Above the cutoff, the
average return was 0.73%.

67
00:02:37.117 --> 00:02:40.521
Only 36 observations from
the sample of 179 sit

68
00:02:40.521 --> 00:02:43.491
above the cutoff,
leaving a smaller group

69
00:02:43.491 --> 00:02:44.577
behind the gap.

70
00:02:45.237 --> 00:02:47.337
In each group, most
returns were positive.

71
00:02:48.690 --> 00:02:51.827
The diagnostics
show 3 eras with 59,

72
00:02:51.827 --> 00:02:53.570
60, 60 observations.

73
00:02:54.590 --> 00:02:56.810
A stored hash
identifies every input.

74
00:02:57.410 --> 00:02:59.542
Together, they make
the study reproducible

75
00:02:59.542 --> 00:03:00.760
and easier to challenge.

76
00:03:01.556 --> 00:03:03.276
Save any study
you want to keep.

77
00:03:03.916 --> 00:03:05.643
A null result is
worth keeping,

78
00:03:05.643 --> 00:03:07.705
because it stops
you testing the same

79
00:03:07.705 --> 00:03:08.596
dead idea twice.

80
00:03:09.504 --> 00:03:11.995
Does the conclusion change
over a quarter instead

81
00:03:11.995 --> 00:03:12.554
of a month?

82
00:03:12.864 --> 00:03:14.614
Keep the predictor
and dates fixed.

83
00:03:15.164 --> 00:03:17.264
Set 63 trading days,
then rerun.

84
00:03:18.318 --> 00:03:20.147
Same predictor, same dates,

85
00:03:20.147 --> 00:03:21.298
a longer outcome.

86
00:03:22.098 --> 00:03:23.848
The correlation is
still near zero.

87
00:03:24.378 --> 00:03:25.762
But looking a quarter ahead

88
00:03:25.762 --> 00:03:27.403
while sampling
monthly means the

89
00:03:27.403 --> 00:03:28.428
windows now overlap.

90
00:03:29.078 --> 00:03:30.752
The count barely moved,

91
00:03:30.752 --> 00:03:31.917
from 179 to 177,

92
00:03:31.917 --> 00:03:35.338
and the observations are
no longer independent.

93
00:03:36.418 --> 00:03:38.518
The saved study keeps
the 1-month version.

94
00:03:41.291 --> 00:03:43.911
9 threshold values were
searched in this run.

95
00:03:44.571 --> 00:03:46.805
The best-looking gap
can be the winner among

96
00:03:46.805 --> 00:03:47.821
several noisy tests.

97
00:03:48.511 --> 00:03:51.131
Only 36 observations sit
above that threshold.

98
00:03:51.991 --> 00:03:54.487
The threshold is also a
percentile of this whole

99
00:03:54.487 --> 00:03:56.671
sample, so it was not
knowable in advance.

100
00:03:57.571 --> 00:03:59.308
Check search breadth,
sample size,

101
00:03:59.308 --> 00:04:01.608
and whether the rule
could have been followed

102
00:04:01.608 --> 00:04:02.221
at the time.

103
00:04:02.658 --> 00:04:05.258
Name a predictor and the
outcome it should forecast.

104
00:04:06.078 --> 00:04:07.397
Read the observation count

105
00:04:07.397 --> 00:04:09.528
before the fitted curve
or best threshold.

106
00:04:10.218 --> 00:04:11.987
Save the study,
change one input,

107
00:04:11.987 --> 00:04:12.898
and run it again.

108
00:04:14.420 --> 00:04:16.660
Choose Take a tour for
a guided walkthrough.

109
00:04:17.220 --> 00:04:18.770
You can dismiss
it at any time.

110
00:04:21.935 --> 00:04:23.400
Open Docs for sampling,

111
00:04:23.400 --> 00:04:24.930
splits, lag, fit method,

112
00:04:24.930 --> 00:04:26.395
and era interpretation.

113
00:04:28.933 --> 00:04:31.144
A signal needs a
real relationship,

114
00:04:31.144 --> 00:04:34.113
enough observations,
and stability across eras.

115
00:04:35.033 --> 00:04:37.400
Without those three,
a smooth curve is still

116
00:04:37.400 --> 00:04:38.153
a null result.