Models with adjustable lookback periods, e.g., the length of a moving average, require a tedious if simple sanity check. Test all periods. Plot the results (final equity value or whatever other measure you are using) vs. period. Which are most profitable? Is a contiguous block of periods profitable?

For example, perhaps only shorter periods are profitable and longer periods lose money. Or are only a few scattered periods profitable? The last effect suggests the model is nonsense. If most periods are profitable, or if all periods between 1-30 days are profitable, that makes some sense. If the 24 and 66-day periods are profitable, but no others are, what real effect could be present to justify their profitability?

Many people use a 200-day moving average to estimate long-term trends. The model hypothesis is that long-term trends matter. I would not consider using a 200-day moving average to be data snooping or data mining. To me, data snooping is testing all possible periods and using the best one without doing any analysis like the one I suggested above. Poor analysis is testing only one period, even if it is your hypothesis, without checking the alternatives at all.


WordPress database error: [Table './dailyspeculations_com_@002d_dailywordpress/wp_comments' is marked as crashed and last (automatic?) repair failed]
SELECT * FROM wp_comments WHERE comment_post_ID = '999' AND comment_approved = '1' ORDER BY comment_date




Speak your mind


Resources & Links