The 5day ROC Volume effect is only short term

24Mar10

Having no success with NN for forecasting next day close prices, I had an idea that I should
train my NN based on an effect that is statistically significant.
Like this effect:
http://cssanalytics.wordpress.com/2010/03/04/adjusting-to-current-market-conditions-the-increasing-importance-of-spy-volume/
It says that when the 5 day ROC (Rate of Change) of the SPY is > 0, the chance of having an Up day next day is 75%.
It looked great. That research looked sensible.
If this effect exists then I can train my NN based only the Volume, or based only the 5dROCVolume data only.
The input would be 1 dimension; and I can debug why my NN doesn’t have predictive power.
Or if I am lucky, my NN will have a predictive power, so this can be one component in my prediction system.

Then calculate the Rate of Change (Volume) formula:
( Volume [today] - Volume [n days ago] ) / Volume [n days ago]

I made a test in Matlab:

for i = 6:1:length(volumes)-1
percentChange = (closes(i + 1)/ closes(i) - 1);
d5ROC = (volumes(i) - volumes(i - nDaysInRoc)) / volumes(i - nDaysInRoc);
isd5ROCUp = d5ROC > 0;
if (isd5ROCUp)
if (percentChange > 0)
nUpAfterVolumeUp = nUpAfterVolumeUp + 1;
pUpAfterVolumeUp = pUpAfterVolumeUp + percentChange;
else
nDownAfterVolumeUp = nDownAfterVolumeUp + 1;
pDownAfterVolumeUp = pDownAfterVolumeUp + percentChange;
end
else
if (percentChange > 0)
nUpAfterVolumeDown = nUpAfterVolumeDown + 1;
pUpAfterVolumeDown = pUpAfterVolumeDown + percentChange;
else
nDownAfterVolumeDown = nDownAfterVolumeDown + 1;
pDownAfterVolumeDown = pDownAfterVolumeDown + percentChange;
end
end
end

I tested it from about 2000-2010. That is about 2500 days, instead of the 60 days tested by Varadi.

Result:
percentUpAfterVolumeUp: 52.76%
percentUpAfterVolumeDown: 51.75%

Not significant difference. Nothing like the 75% Win% that is reported by Varadi.

avgpUpPercentAfterVolumeUp: 0.91%
avgpUpPercentAfterVolumeDown: 0.86%

So, after 5dROCVolume was up, and we have an Up day next day, the %percent gain tends to be greater.

avgpDownAfterVolumeUp: -1.04%
avgpDownAfterVolumeDown: -0.95%

So, after 5dROCVolume was up, and we have an Down day next day, the %percent loss tends to be greater.

So, a simple strategy that would play Long after UpVolume day and play short after DownVolume day would be doomed.
😦

Later, I was wondering whether the test made by Varadi was proper or not.
I think, he is good, but his main mistake is that he considers only 3 months, = 60 days in the test.
The 75% Win% probability inspected after 60 samples is just a too insignificant thing to draw conclusions.

To be candid, Varadi only says that this is the effect in the last 60 days, he says nothing about its long term applicability. (albeit the reader of the blog very easily falls into the trap and makes this conclusion. Varadi made the backtest. He is surely aware that it fails in the long term. An honest researcher should emphasize not only the pro but also the con of his test.)
Varadi says that in ‘this climate’ (bullish), this is what works.

So, there can be 2 reasons for Varadi’s lucky 75% Win% finding (that is not validated by me) and that in the long term this is not a prediction:
1.
The effect doesn’t work.
He only was simple lucky, because, he used only 60 samples. Fool’s gold.
2.
The effect really works, but it worked only in that 60 days period.

Suggestions for future research:
1.
One would argue for a strategy that has a 60 days rolling window.
In that window, we can make statistical measurements, and we can play according to this statistics the next day.
This strategy may worth a test, but it needs more programming.

2.
Another idea that we may leave the 1 day forecast and concentrate on mid-term, like weekly SPY forecast.
That is not so competitive place (not so many researchers investigate it) and there are less randomness in the weekly
SPY prices and more dependence of fundamental factors like FED interest rates, option expirations, etc.

Advertisements


No Responses Yet to “The 5day ROC Volume effect is only short term”

  1. Leave a Comment

Leave a Reply

Fill in your details below or click an icon to log in:

WordPress.com Logo

You are commenting using your WordPress.com account. Log Out / Change )

Twitter picture

You are commenting using your Twitter account. Log Out / Change )

Facebook photo

You are commenting using your Facebook account. Log Out / Change )

Google+ photo

You are commenting using your Google+ account. Log Out / Change )

Connecting to %s


%d bloggers like this: