Hi Dhalsim, The system is intra-day only. I know the entries have a positive expectancy coupled with the exits, the large stop is due to the fact the system takes 1 shot at the setup ... sure, I could make things way more complex, using small stops & multiple attempts, at the end of the day there will be more comms, more slippage, and of course a lower win%. If you like a reasonable stop, use 1pt with this system, and get 85% of the system's potential P&L (but increased max DD). No, I don't find it frustrating to trade CL with a real edge and large stops ... sure, the P&L curve has larger ups & downs than with tiny stops / targets, but I find that reducing the % taken by comms & slippage (through reducing the number of trades) is a serious advantage. Of course, if I had the recipe for a profitable trading system with 10-tick initial stop & 10-ticks average net per trade after comms & slippage, I would showcase it. (I would showcase it even with 5 or 3 ticks net per trade, after comms & slippage). But I doubt this is possible for an automated trading system. Cheers D.
This pattern is a candidate to be added to that new system now named "CL Selective". A quarter of its P&L comes from 2008 alone (how surprising for a short setup on CL), that being said most of the other years have good performance, aside from 2003/2004 which are just about BE. I would discard the pattern for its too small base, if it wasn't for the fact its frequency, win% & avg/trade are kinda consistent across the years. Comments?
It was a productive week overall on the strategy development front. The current (semi-final) version of CL Selective uses 3 Long patterns + 4 Short patterns. I decided to use 150-ticks catastrophic stops, this is a little larger than I would like (I use 120-ticks for CL Always-In) but the overall penalty at 150-ticks is 10% of the raw P&L (no-catastrophic stop), and gets close to 15% at 120-ticks. The per-trade performance is much better than CL Always-In, but the patterns have a much thinner base (508/792/163 for the Longs, 428/267/542/152 for the Shorts) (some Short patterns overlap), with much higher risks of over-fit to past data. For those worried about the initial stop, the system is profitable from 10-ticks initial stop, although with a ridiculous performance. It gets better fast: - 10-ticks: P/F 1.21, win% 13.5%, avg/trade $23, maxDD : -11200 ; total P&L 10% of "no-stop" - 20-ticks: P/F 1.43, win% 26.2%, avg/trade $71, maxDD: -9800 ; total P&L 32% of "no-stop" - 30-ticks: P/F 1.54, win% 35.8%, avg/ trade $109, maxDD: -7300 ; total P&L 50% of "no-stop" but from there, it is a slow grind to capture the other 50% of the available P&L. - 50-ticks: P/F 1.57, win% 47.2%, avg/ trade $139, maxDD: -9300 ; total P&L 63% of "no-stop" - 60-ticks: P/F 1.57, win% 50.2%, avg/ trade $146, maxDD: -10800 ; total P&L 66% of "no-stop" - 70-ticks: P/F 1.58, win% 52.6%, avg/ trade $154, maxDD: -11300 ; total P&L 70% of "no-stop" - 80-ticks: P/F 1.58, win% 54.3%, avg/ trade $158, maxDD: -12900 ; total P&L 72% of "no-stop" - 90-ticks: P/F 1.61, win% 55.5%, avg/ trade $168, maxDD: -13900 ; total P&L 76% of "no-stop" - 100-ticks: P/F 1.65, win% 56.7%, avg/ trade $178, maxDD: -11700 ; total P&L 81% of "no-stop" - 110-ticks: P/F 1.67, win% 57.4%, avg/ trade $183, maxDD: -10800 ; total P&L 83% of "no-stop" - 120-ticks: P/F 1.71, win% 58.1%, avg/ trade $192, maxDD: -11100 ; total P&L 87% of "no-stop" ... - 150-ticks: P/F 1.73, win% 58.8%, avg/ trade $198, maxDD: -11400 ; total P&L 90% of "no-stop"
So would you say essentially you are running 7 different strategies all tied together on this one symbol. Can you only trade one pattern at a time? Also, do all the patterns therefore have same set stop??
It is a matter of semantics ... to me it is one strategy (the tools & techniques used to identify & trade those patterns), each pattern is a component of that strategy, but from the outside it appears like trading 7 different strategies. Of course yes. The stop is a "catastrophic stop" which isn't meant to improve the system performance, but to limit the risk when the market throws at us a long series of large loosing trades. So I use the same value for all patterns, the one adjustment that I am willing to explore again (*) is to base that catastrophic stop on the current volatility (using a 60 to 250-days ATR). (*) I had done that research for CL AlwaysIn, and it didn't do any better than a fixed size catastrophic stop.
Have you tried testing the 7 patterns individually with different stops and targets? If they are 7 different patterns which you have backtested results for then i would think it would be best to have different stops and targets for each pattern. I presume each one of the patterns is working on a trailing stop based on what you have said. Hence, the fact you only need to catastrophic stop before a trail gets kicked into place. I like the results you have posted but i don't like the fact that your strategy has seven different types of patterns it trades of. To me that is 7 different strategies. The problem i see with you strat is that over 10 year period for intraday strategies it has too few trades. I would say the one with 792 trades over 10 years is okay but anything less than that is too small a sample and too easy to curve fit. I try to average 100 trades per year on my intraday 1min chart strats. Another problem is the fact that the initial stop affects the profit factor so much. Also, it is a set stop of fixed amount which is always considered curve fitted. Imo stop should be dynamic and never static fixed amount. Profit factor to me is also too low for mixing seven strategies. I try to aim for PF 1.9 or above when combining different strats on same instrument. Apart from that the Sharpe and draw-down look good. Performs well in volatility it seems as should any strategy as volatility is where us traders should make most of our money. Also, anyone reading my post take it with a pinch of salt and consider that i am not an expert but a newbie to automated trading who has only taken his strats live for last 3 months. So take my advice for what its worth.
Hi Dahlsim, Thanks for the feedback. I am certainly conscious of the risks of over-fit to past data, and I realize that adding over-fit patterns is a way to speed-up the path to failure. In that spirit, I am not willing to customize the trading management on a per pattern basis. To me, that would be typical over-fit. The smaller base patterns have some reasons to be smaller base (eg, if you have a trade setup for FOMC meeting, you won't see it more than 6 or 7 times a year ... or one setup for the Crude Inventory Report will only be in effect once a week). But that doesn't deny the possibility of over-fit to pseudo-random outcomes. Thanks again D.
DOM, try volatility adjusted stop with a shorter lookback. I once tested 'predicted' volatility vs realized volatility (ATR) and it came out that up to using past 10 days of data the variance of 'prediction minus reality' was decreasing but not anything beyond that.
CLSelective v09 backtest performance data, using 150-tick initial stop I still have to test an ATR-based initial stop.
CLSelective v10 backtest performance, 150-ticks initial stop I am done with the performance side of this system (at least, in the short-term), I also have added a couple of operations-related features, should be all done next week. Real-time performance over the last 4 weeks is showing 26-ticks of positive slippage (vs backtest) for 15 trades fully automated, which is a very good sign (although 15 trades quite a small sample set).