Yes, I used "ranges" in the context of time frame, e.g. if you have 10 years of data, test 1 would use the first 5 years and test 2 the second 5 years. This can be significantly different than doing 10 years of testing in a single block when you consider drawdown and extreme trades and average trades in the separate intervals. The presumption is that the "curve" to be fitted is always changing and that what exists in one interval may or may not exist in another. It's mostly directed at evaluating the effectiveness of parameters or optimization if that's been used. Using lots of stocks adds to your statistical database, but you may not get trades on all stocks. Your criteria, for example, of using appropriately priced stocks will eliminate many samples from consideration and will eliminate certain intervals on samples that do qualify.
The reason this came up is that a couple years ago Richard commented that he'd never gotten any system to test well on IBM. I wondered if getting something to work with IBM would provide a system that worked better on other stocks (it doesn't), so I went to work on the most recent 3 years of data. But when I looked at the previous 3 years I had a different mix of results. When accumulated in a single summary, the results were acceptable, but when I looked at period 1 vs. period 2, there was no correlation. I noticed similar results on other stocks. Ideally what you'd like to see is consistency across the range of data.
re: best-choice indicators/parameters... I haven't tested the idea, either, of identifying characteristics of price movement. I had a notion that there's something done in radar signature analysis that could be applied, but I'm not informed enough to know how to approach the design. I even had the offer of help from someone that does that sort of thing, but I didn't even know what to ask. I suspect that there are statistical tests that could be applied as well, but again I'm out of my league. So this is pure speculation and I always hope that someone will someday say "that's a piece of cake!" and presto! we'll have a money maker.
re: curve fitting... your conclusion is what virtually everyone does in one form or another since we can't pick and choose indicators for the current data characteristics. My only point is that the best profits come from trades that happen when the indicators and the price data correlate, i.e. the curve is fitted. Success without automated adaptation is limited to a balance of acceptable:poor results, hopefully biased toward acceptable. Screening (filtering) is possibly your best tool.
re: splits... using price screens is your best choice, but it limits your test span unless the stock trades in a range. You can get around that by resetting your screen and going thru the process again, taking into account the smaller or larger price and what trade size would scale properly to "real" perferred prices. E.g., $10 stock and a $10,000 trade scale to a $20 stock and a $20,000 trade or a $5 stock and a $5,000 trade. Another approach is to use multiples of price for establishing the size of a trade, e.g. if price = 10 and the multiple is 1000, then your trade size is 10,000; for a 12.750 stock you'd use the same multiple and the trade size would be 12,750. This is just another way of saying the same thing as "scaling", but it translates into a codeable system instead of being a screen. The problem is that you have to realize that you may be looking at trades that you won't take, but at least you can summarize the results with percentages and have some confidence you're comparing similar arithmetic groupings.
Bruce |
