In our historical simulation, there were 17 losing trades and 10 winning trades. The first ten trades ran in the following order: winner, loser, loser, winner, loser, loser, loser, winner, loser, loser. I can type out the order for the entire 27 trades in the system, but you get the idea.
What's to say the order of wins and losses will be exactly the same in the future? Answer: nothing. In fact, we'd be very surprised if the next 27 trades followed the same order as the first 27 trades. The fact that the first 27 trades fell as they did is completely the result of luck. And the equity curve derived from this order is also the result of luck. If we are to derive a reasonable expectation of what our drawdown could be going forward, we need to take out the element of luck. How do we do that? We randomize the trade order. Enter Monte Carlo simulation.
Think of this as shuffling cards, and taking a snapshot after each shuffle. All black cards are winners and all red cards are losers. We have the deck stacked with 17 red cards and 10 black cards. For those with statistical savvy, you'll recognize there is not even a valid sample size of 30 to draw statistical inference from in our exercise. At this point, we're just going through the process, and make no claims about the validity of the 50/200 cross as a tradeable system.
Since I use TradersStudio, the following is Monte Carlo data derived from the software. I'm also using a Monte Carlo Test macro designed by Rich Denning, who was kind to contribute the macro in a recent issue of TASC.
First, let's take a look at the original equity curve for the system:
After running the Denning Monte Carlo macro, we have the following distribution chart. The Y axis is the total number of trades sequences that have fallen into the X axis bin. The X axis is the level of drawdown. You will notice that most of the trade sequences (about 108) had a drawdown of $99,760. The total number of trade sequences run were 1,000.
Finally, here is a terminal report that shows what the drawdown expectation is given different confidence levels:
Since our distribution is not normal, we have limited utility for standard deviation in determining confidence levels. Instead, we take the percent of simulations to find the interval. This takes into account the kurtosis natured skew of our distribution.
What's interesting is that our historic simulation's max drawdown of $81,000 is quite a bit below the average of $104,000 and the median of $128,000. In fact, we really shouldn't get too excited about the drawdown on this system until it exceeds the 95% level, which is $149,000.
As stated earlier, these statistics are more or less a practice run of the process and a definition of best practices in systems development. It's clearly almost ridiculous to draw any statistical conclusions from a system that averages about one trade per year.
3 comments: