"Causal Analytics"
is ...
... the study of cause and effect relationships found in data through
inductive reasoning. In the case of BioComp Profit, it is the study
of which input variables cause your target variable to move based on your
data. Causal Analytics seeks to answer
three important questions:
-
Does Indicator A cause
My Target to Move?
-
When?
-
By How Much?
|
|
Time
is Important in Market Timing
We theorize that cause and effect relationships are
manifested in TIME and there is always a lapse in
time, regardless of how large or small, between causes and effects. If A drives B, and A changes, then B
will change after some amount of time elapses, whether that is picoseconds
or eons. Time tells us that If B occurs before A, then A cannot be a
cause of B.
Thus it is helpful to sample data as fast as the causal time delay to see
the causal relationships. In Profit, however, we are working with
end of day data. Because of this, we cannot see OR USE the intra-day
effects. If something happens during the day to increase the price
of our security, we cannot see the cause then the effect in the end of day
data, we just see the numbers moving together, nor can we capitalize upon
it, because the move has been made and it was unforeseeable
beforehand. Intra-day systems have an advantage here, but at other
disadvantages, which is a separate topic.
So, what we are looking for
are indicators that lead by at least 1-2 days, maybe more.
|
|
The Importance
Causal Analytics is important in
market timing systems because it helps us...
-
Determine if our
indicators are useful or not. Since the markets are somewhat
of a mystery, it's difficult to know which indicators are useful for
forecasting. We spend a LOT of time in trial-and-error 'runs'
trying to sort out the good indicator from the worthless. We
generally make improvements in our systems as we go but we know the
old saying "Garbage In, Garbage Out". That applies to
indicators we THINK are good, but are not. If we use a
non-useful indicator in our models, we will do more damage than good
because our signals will change with the useless indicator's changes
and thus the signals are changing for no reason.
-
Eliminate
Cross-Correlated Indicators. The indicators we are using are
likely to contain the SAME information with a few differences.
For example, the NYSE is highly correlated with the SP500
futures. This is often called "cross-correlation" or
"co-linearity". This problem leads people to try to
use Principle Components Analysis (PCA), which attempts to reduce
co-linearity and remaps our variables into some weird coordinate
system that only guys in lab coats understand. Profit users
report mixed results with PCA. Causal Analytics attempts to
resolve this by finding the most causal variables in sequence that
explain residual error in signals (predicted vs. actual signal).
This gives us a list of key variables that are least cross-correlated
and most causal of our signals. (This capability will likely be in the
next major release of BioComp Profit, version 7.0).
-
See the shifting causes in
the non-stationary markets and determine when we are financially
" in and out of
control". By trending causality in a window, we can see
our key variables gain and lose causality in the markets, giving us
more information than just watching equity curves rise and fall.
This feature will come in a later edition of Profit.
Thus, Causal Analytics is
important to finding true leading indicators, the utility of the ones we
have to build better models and forecasts and better understand the
dynamics of the markets.
|
