For what it's worth. I did some analysis and discovered some other useful criteria. For example, entering on a "hook day" also has good results: if ( today's open > yesterday's high + x ) or ( today's open < yesterday's low - x ) then trade today I found a good value of x of around 3. Richard
i'm counting something like 10-15 degrees of freedom so far in this 40yroldtrader system and it sounds like there are many more undisclosed. just because i've never seen it obviously says nothing, but i've never seen something functional with anywhere close to that many moving parts. i definitely start seeing curve fits over about 5 degrees, depending on what types
Degrees of freedom are reserved for use with normally distributed data. I have tested enough market data to know the data are not normally distributed and are non-stationary. To that end if a test is done on 1,000 days of data, you would have to assume 1,000 degrees of freedom based on the data alone. I have found no value for degrees of freedom, significance testing, or sampling errors because they don't apply to the markets. If you disagree I'd love to hear how you fit non-normal non-stationary data into a normal distribution so that t-Tests could be performed and have some relevant value.
on a quick analysis basis, your profit factor of 1.46 is lousy, as is your profit/drawdown ratio (rough sterling ratio). do only way to see if you have a robust system is to post results, applied to the other indexes, with no tweaking. if the results on another index is roughly 70% of these results, then you have a robust system. remember no tweaking!!!!! Paula
i probably misused the term. i simply meant to point out that systems are usually curve fit with that many inputs/vars. it's observable with optimization on one set of history and application to another.
Here's something interesting. I tried to do the edge test that acrary described in his posts. I simulated random trades using the same holding period. Then I ranked my total pnl against the total pnl from each set of random trades. The one difference is that most of my trades are overnight, so I did not exit on close like acary did. However, the principle is the same in that I kept the holding period the same. Here are the results. 2006: 30% 2005: 88% 2004: 99% 2003: 93% 2002: 90% 2001: 98% 2000: 86% 1999: 43% 1) Do the numbers from 2000-2005 look obviously wrong to you? This is my first time with the edge test, so I don't have perspective. However, I did check my numbers on a test run and it did seem to be calculating correctly. The test run had far fewer trades and trials though. 2) This strategy was designed using data from 1998 to 2005. The edge (if it had one) and profitably dropped off significantly in 2006. Do you think this means I curve fitted this? 3) How would this edge test catch a curve fit system? I had thought that if the system was curve fit, then it would not rank very high in the edge test.