WEBVTT

1
00:00:00.306 --> 00:00:02.876
The market keeps
setting record highs.

2
00:00:03.252 --> 00:00:04.620
Should you wait for a dip

3
00:00:04.620 --> 00:00:05.441
before you buy?

4
00:00:06.886 --> 00:00:08.374
The study needs a target

5
00:00:08.374 --> 00:00:09.552
and a fixed window.

6
00:00:09.962 --> 00:00:12.438
The target is the
same broad stock fund.

7
00:00:12.931 --> 00:00:17.216
15 years, from
2010 through 2024.

8
00:00:17.769 --> 00:00:19.297
Fixing the dates
is what makes

9
00:00:19.297 --> 00:00:20.519
the result reproducible.

10
00:00:20.855 --> 00:00:23.406
Signal Analyzer compares
a predictor with the

11
00:00:23.406 --> 00:00:25.617
outcome that followed
each observation.

12
00:00:26.067 --> 00:00:28.023
It measures
relationship strength

13
00:00:28.023 --> 00:00:30.156
and checks whether
that relationship

14
00:00:30.156 --> 00:00:31.401
persists across eras.

15
00:00:31.901 --> 00:00:34.095
As an example, we will
test whether a broad

16
00:00:34.095 --> 00:00:35.779
stock fund's price
level predicts

17
00:00:35.779 --> 00:00:36.901
its next-month return.

18
00:00:38.188 --> 00:00:40.381
The predictor is the
fund's price level.

19
00:00:40.836 --> 00:00:43.778
Another study could
use a moving average,

20
00:00:43.778 --> 00:00:45.716
a trailing return, a yield,

21
00:00:45.716 --> 00:00:47.223
or a calendar effect.

22
00:00:47.671 --> 00:00:49.501
A predictor can
also combine several

23
00:00:49.501 --> 00:00:50.771
of those transformations.

24
00:00:52.182 --> 00:00:54.485
The outcome it has to
predict is the return

25
00:00:54.485 --> 00:00:56.091
over the next
21 trading days.

26
00:00:56.483 --> 00:00:57.913
Roughly one month ahead.

27
00:01:00.050 --> 00:01:02.486
Sampling and splits
decide what counts

28
00:01:02.486 --> 00:01:03.576
as evidence here.

29
00:01:04.045 --> 00:01:06.952
Spacing decides whether
two observations overlap,

30
00:01:06.952 --> 00:01:08.851
and overlapping
observations are

31
00:01:08.851 --> 00:01:10.334
not independent evidence.

32
00:01:10.802 --> 00:01:13.827
Predictor lag decides
how much of the past the

33
00:01:13.827 --> 00:01:15.471
signal is allowed to see.

34
00:01:15.942 --> 00:01:18.169
The monthly samples
are divided into

35
00:01:18.169 --> 00:01:19.468
3 historical periods.

36
00:01:21.047 --> 00:01:23.619
Fit settings control how
the relationship is drawn.

37
00:01:24.071 --> 00:01:26.127
Choose a smooth
polynomial curve,

38
00:01:26.127 --> 00:01:28.930
or bins showing the
average outcome by group.

39
00:01:29.383 --> 00:01:31.892
This study uses the
default smooth curve,

40
00:01:31.892 --> 00:01:33.544
which is worth remembering:

41
00:01:33.544 --> 00:01:35.502
a curve can look
convincing even

42
00:01:35.502 --> 00:01:37.276
when the
correlation is zero.

43
00:01:38.702 --> 00:01:41.152
Does price level predict
the next month's return?

44
00:01:41.348 --> 00:01:43.921
Run 15 years of
monthly samples with

45
00:01:43.921 --> 00:01:46.208
non-overlapping
forward windows.

46
00:01:46.726 --> 00:01:49.010
First, understand
the threshold sweep.

47
00:01:49.435 --> 00:01:51.577
Each cutoff splits
the same sample

48
00:01:51.577 --> 00:01:53.279
and compares outcomes above

49
00:01:53.279 --> 00:01:54.098
and below it.

50
00:01:54.493 --> 00:01:56.751
The same observations
are tested against

51
00:01:56.751 --> 00:01:57.824
9 threshold values.

52
00:01:58.259 --> 00:02:01.073
Repeated tests can
produce an attractive gap

53
00:02:01.073 --> 00:02:02.160
from noise alone.

54
00:02:02.646 --> 00:02:04.035
Read the observation count

55
00:02:04.035 --> 00:02:05.211
before the widest gap.

56
00:02:06.118 --> 00:02:09.253
179 independent
observations.

57
00:02:09.613 --> 00:02:11.587
The correlation
between price level

58
00:02:11.587 --> 00:02:13.674
and the next month's
return is 0.000,

59
00:02:13.674 --> 00:02:15.028
to three decimal places.

60
00:02:16.011 --> 00:02:18.199
On the chart that
is a flat line.

61
00:02:18.632 --> 00:02:21.976
Where the market sits tells
you almost nothing about

62
00:02:21.976 --> 00:02:23.198
where it goes next.

63
00:02:25.177 --> 00:02:27.559
The sweep is where
research goes wrong.

64
00:02:28.047 --> 00:02:31.790
Below the cutoff at 80%
of the price distribution,

65
00:02:31.790 --> 00:02:34.336
the forward return
averaged 1.28%.

66
00:02:34.804 --> 00:02:38.996
Above the cutoff, the
average return was 0.73%.

67
00:02:39.496 --> 00:02:42.844
Only 36 observations from
the sample of 179 sit

68
00:02:42.844 --> 00:02:45.765
above the cutoff,
leaving a smaller group

69
00:02:45.765 --> 00:02:46.833
behind the gap.

70
00:02:47.323 --> 00:02:49.896
In each group, most
returns were positive.

71
00:02:51.112 --> 00:02:54.909
The diagnostics
show 3 eras with 59,

72
00:02:54.909 --> 00:02:57.019
60, 60 observations.

73
00:02:57.491 --> 00:03:00.349
A stored hash
identifies every input.

74
00:03:00.844 --> 00:03:03.754
Together, they make
the study reproducible

75
00:03:03.754 --> 00:03:05.417
and easier to challenge.

76
00:03:05.931 --> 00:03:07.531
Save any study
you want to keep.

77
00:03:07.960 --> 00:03:09.859
A null result is
worth keeping,

78
00:03:09.859 --> 00:03:12.127
because it stops
you testing the same

79
00:03:12.127 --> 00:03:13.107
dead idea twice.

80
00:03:13.638 --> 00:03:16.518
Does the conclusion change
over a quarter instead

81
00:03:16.518 --> 00:03:17.165
of a month?

82
00:03:17.464 --> 00:03:19.466
Keep the predictor
and dates fixed.

83
00:03:19.962 --> 00:03:22.917
Set 63 trading days,
then rerun.

84
00:03:23.820 --> 00:03:25.983
Same predictor, same dates,

85
00:03:25.983 --> 00:03:27.345
a longer outcome.

86
00:03:27.719 --> 00:03:29.815
The correlation is
still near zero.

87
00:03:30.290 --> 00:03:32.081
But looking a quarter ahead

88
00:03:32.081 --> 00:03:34.203
while sampling
monthly means the

89
00:03:34.203 --> 00:03:35.530
windows now overlap.

90
00:03:35.901 --> 00:03:37.914
The count barely moved,

91
00:03:37.914 --> 00:03:39.314
from 179 to 177,

92
00:03:39.314 --> 00:03:43.427
and the observations are
no longer independent.

93
00:03:43.876 --> 00:03:46.258
The saved analysis keeps
the 1-month version.

94
00:03:48.937 --> 00:03:51.413
9 threshold values were
searched in this run.

95
00:03:51.884 --> 00:03:54.895
The best-looking gap
can be the winner among

96
00:03:54.895 --> 00:03:56.264
several noisy tests.

97
00:03:56.722 --> 00:03:59.864
Only 36 observations sit
above that threshold.

98
00:04:00.266 --> 00:04:03.160
The threshold is also a
percentile of this whole

99
00:04:03.160 --> 00:04:05.693
sample, so it was not
knowable in advance.

100
00:04:06.173 --> 00:04:07.910
Check search breadth,
sample size,

101
00:04:07.910 --> 00:04:10.210
and whether the rule
could have been followed

102
00:04:10.210 --> 00:04:10.823
at the time.

103
00:04:10.925 --> 00:04:13.525
Name a predictor and the
outcome it should forecast.

104
00:04:13.947 --> 00:04:15.440
Read the observation count

105
00:04:15.440 --> 00:04:17.851
before the fitted curve
or best threshold.

106
00:04:18.263 --> 00:04:20.589
Save the study,
change one input,

107
00:04:20.589 --> 00:04:21.787
and run it again.

108
00:04:23.251 --> 00:04:25.829
Open Help,
then choose Take a tour for

109
00:04:25.829 --> 00:04:27.253
a guided walkthrough.

110
00:04:27.645 --> 00:04:29.195
You can dismiss
it at any time.

111
00:04:34.469 --> 00:04:37.728
The Signal Analyzer
guide covers sampling,

112
00:04:37.728 --> 00:04:39.591
splits, lag, fit method,

113
00:04:39.591 --> 00:04:41.531
and error interpretation.

114
00:04:42.929 --> 00:04:45.326
A signal needs a
real relationship,

115
00:04:45.326 --> 00:04:48.546
enough observations,
and stability across eras.

116
00:04:48.984 --> 00:04:51.873
Without those three,
a smooth curve is still

117
00:04:51.873 --> 00:04:52.792
a null result.