#018: HOW I SPENT TOO MUCH TIME WITH 3D OPTIMIZATION CHARTS
Some traders ask me if I use 3D optimization charts for my own trading. Do I have good experience with this tool? Would I recommend it for robust ATS development? The answers are quite simple: NO!
I have spent a lot of time with it, so I can explain it more in detail (luckily I have never actually started using it for my ATS development).
First, let's make a small recapitulation - what exactly is the 3D optimization chart? It is a simple visual aid that helps us to find global optimal values instead of just local ones.
The local optimum is a combination of 2 parameters that gives us the best results within a specified range, no matter how well surrounding parameters work.
The global optimum is a combination of 2 parameters that gives us positive results within a particular parameters combination of both variables, but gives us also stable and quality results for all surrounding parameter combinations. The more combination surrounding the “central” parameter set work, the better.
The main advantage is that the whole concept seems to be simple, logic and robust and is practically easy to grasp, with a nice visual interpretation.
On the other side, the biggest disadvantage is that you can use this concept only for optimization of 2 parameters (there are certain tweaks how to use this concept also for other parameters, but then its basic simpleness fades away).
Let's take a look at a simple example - a simple model of a breakout strategy which uses the close price of x-bars back increased by 3x ATR with a certain period as a starting point. Exit at the end of the day.
The simplified model can look like this:
Buy next bar at ( Close[BarsBack] + 3 * ATR(ATRperiod) ) Stop; Setexitonclose;
We want to see what are the optimal parameters for BarsBack and ATRperiod parameters. For a start, we use the range 1-10 for both parameters (in reality, we would use another range, but for the sake of this example let’s use 1-10 values) and create a simple, easily visually interpreted 3D chart that will display NET PROFIT for all combinations.
As you can see on the chart, the best values from this range would be 10 (ATRperiod) - 1 (Barsback) combination. We can consider this combination to be the local optimum - and, most probably, it can be just a result of over-optimization.
On the other side, the 5-3 combination looks more like the global optimum, as also the surrounding combinations have similar results as the 5-3 combination. We skip the point that the global optimum should be much more extensive; we use this as a simple example.
Until this point everything is alright.
So where is the problem?
The problem is high instability of this approach.
As a full-time ATS trader, I spend a lot of time comparing different approaches and methods. I have also spent a lot of time with 3D optimization. I have found out, as well as other users, quite tempting the simplicity of this concept. Roughly a year ago I performed a really extensive series of tests that took me one month of really intensive work. The reason behind all this work was to find global optimum parameters for a lot of breakout models in different time slots using 3D optimization chart.
BREAKOUT MODEL A: Searching global optima in 2003–2005, 2006–2008, 2009–2011, 2012–2014.
BREAKOUT MODEL B: Searching global optima in 2005, 2006, 2007, 2008, 2009, 2010, 2011, 2012, 2013, 2014.
The objective of the test was to determine how much the global optima change during each test period.
The conclusions were shocking. From previous, less intensive tests, I knew that the parameters obtained this way are not stable. However, this through test showed me that the instability can be enormous. For example - in many cases the best global optimum for one time slot was almost the worst possible to the next one. The global optima has incredibly changed - if I use the 3D chart mentioned above, for one year it would recommend the global optima as 9-9 and for the next year 2-2 (in the opposite corner).
This instability, tested across a range of markets, breakout models and time-frames, definitely reassured me that the 3D optimization charts method is not for me. Therefore, I am not using it for my own trading at all.
Of course, for some one you the 3D optimization can work better (remember, there is no dogma in trading), but this concept is not for me.