I only have 1 parameter that I optimize on, it is the decision threshold (as each trading decision is taken from a score, which is a linear combination of each applicable pattern (avg & info.ratio), which values are directly out of stat analysis). Funniest thing is, I had forgotten to run that optimization on this version (well, I didn't work on this system for about 6 months). The final result for v50 is marginally better than what I posted (603k vs 598k with a few tens less trades). Re. the patterns themselves, I use 6 "metrics" across all patterns (most use 2, one use 1, one use 3). To clarify further re. optimization vs stat. analysis, all patterns are defined based on the statistical analysis of 2007..2012. They are "optimized" at that level (meaning, for each metric I select the range that provides the best result for *that* pattern in stat. analysis). Patterns which use 2 or more metrics, in general have only 1 "flexible" metric (the other(s) are used to define the global context)
I'll show a couple example of *stand-alone" patterns, always-in. Pattern 141: (2280 samples in stat analysis)
Patterns 141+142 (total 4200 sample in stat analysis) Used to be a single pattern in v45, I did split it in 2 parts in v46, pattern 142 has lower score than 141 & in practice is never triggered alone (its individual score is way below the decision threshold)
The one thing which is "spectacular" with 141 & 151, is that they can be run as the only pattern, always-in, and are profitable just by themselves. All other patterns are supplemental - each contributes to improve the system's P&L, but none could be run by itself. Pattern 13x (technically 131+132+133) is my DF20 reversal pattern. The performance in always-in mode clearly shows this pattern is no longer the dominant reversal pattern.
Results for the week ending March 211st, 2014: - 6 win ; 7 losses ; 1 BE ; net -2220 of which -200 result of a human error. This week marked the 1st year of trading live CLAlwaysIn (on Wednesday, March 19th) ... I thought I would give a few highlights, good & not so good: - 598 trades in 1 year, in the market about 100% of the time (except for those rare circumstances of being stopped & waiting for the next trading signal), all trades & rollover executed 100% automatically. - 2 datafeed issues required manual intervention (all others were correctly handled automatically by the software ... this is 3 reloads every night as the IB servers reset, plus on average 1 reload / day upon network issues) - 2 historical data issues rooted in software ... as of February 2014 the corresponding subsystem is as bulletproof as possible, it does detect loss of connectivity happening *during* the historical data download, and as a results the strategies go idle when restarted on possibly incorrect data while the system does another historical data download - 1 human error, believe it or not, on the exact 1 year anniversary of trading that system. - Peak P&L: +21130 on July 3rd - Max DD: -22495 (from P&L peak) on October 15 - P&L at last trade closed March 18th: +8770 (for 1 contract traded) corresponding to +25% on the capital at risk - Current P/F: 1.05 - Current Win%: 52% Even though the financial performance is nowhere close to my expectations, this is nonetheless a positive result, with a lot of progress made along the way: - the infrastructure software has demonstrated its ability to run 24/7 & support an "always-in" strategy - the financial performance, although meager, demonstrate (IMO) the viability of the underlying concept, and the existence of one or several edges in this system, - the versions since last summer use a purely algorithmic mechanism to take all valid patterns in the trade decision process, each pattern-rating values directly extracted from the statistical analysis - the versions since October have a reduced set of patterns, in an attempt to reduce possible over-fit - I now use statistics generated using the underlying market model to assess each pattern vs random entries (more on this later) Another interesting outcome is the creation of the CL151 system, which is based on pattern 151, the most stable & second largest pattern of CLAlwaysIn - this single-pattern system, is my best system (IMO) to date (based on the pattern sample size, raw performance - ie., performance before any trade management & trade filters, and separation vs random entries - both before & after filters). The v51 of CLAlwaysIn (to go live Sunday evening) brings significant perfomance improvement in backtest (about +20%) - through refining one pattern (151, thanks to the work done for CL151), and the addition of 2 minor patterns.
As I was reflecting on this year of trading CL ALwaysIn, and in the midst of developing v51 - still doing stat analysis on 2007..2012, and "forward testing" on 2013, I was puzzled by the somewhat low P&L result for 2013, no matter what I tried. At some point, I realized that overall, 2013 win% & avg/trade was comparable or better than the prior 4 years, and that it was really the number of trades that was down significantly from those years (even lower than 2007). So, I decided to take a look at some stats directly out of the underlying market model. The market model revolves around a trend-change mechanism, and the basic idea behind CL AlwaysIn is to find (statistically) which trend-changes should be traded "with-trend", and which trend-change should be faded. The first aspect that I looked at, was the number of trend-changes. And of course, what I found is that 2013 had a significantly lower number of trend-changes, about 22% less than the average of the prior years (2008 excluded).