The mid-1990s was a golden era for the trading system business. System vendors such as Keith Fitschen, Randy Stucky and Mike Barna thrived. Most of these systems were either trend-following, index, volatility breakout and variations of pivot points.
Hundreds of copies of a trading system were sold during this time. Some developers traded their own systems and made money. Developers learned that those who purchased their systems often ran from adversity, even en route to greater profits, and would stop trading a system amid its first drawdown. The business model was relatively simple. The systems were sold as software. Clients would buy the system, trade it until that first drawdown and then stop. A typical price for a trading system was $1,000 to $2,000.
Vendors could sell these systems for this knowing that only10% or so of the clients would continue trading after a drawdown. This allowed the vendor to continue trading the system even in small markets without concern of saturation. If 100 systems costing $2,000 each were sold, the vendor made $200,000; but because only 10 people were likely to stick with the system, the inefficiencies it tapped into remained. Even if in the best of times they sold 1,000 copies, grossing $2 million, 100 people trading 200-300 lots would have a minimal impact, if any, on a liquid market such as stock indexes.
Despite that secrets are held dear in the trading industry, most vendors sold fully disclosed systems. This became a key to success. If clients didn’t understand a trading system or have access to its logic, it became very easy to pull the plug the moment the system stopped working. Today, most systems are leased with undisclosed logic. Although this is cheaper for the customer, those trading the systems become the weak hands of the market. Folks are quick to abandon a losing position.
Both drawdown and the largest losing trade—key factors many people use to determine whether or not a system is worth trading—are poor predictors of a system’s reliability. In general, the lower the drawdown or the largest the losing trade, the less likely the system is to work in the future.
When developing a trading strategy, it’s important to focus on the core logic, and not just performance metrics. Simple Harmony, which I developed in 2005, was a basket trend-following system that has been ranked in the top 10 by independent system tracker Futures Truth since release.
A modified version, called Trend Harmony, had about 40% less drawdowns than Simple Harmony because of filters. For example, mechanical logic was employed to filter out trades during the fourth wave of an Elliott Wave impulse sequence. Another version traded on multiple time frames. Both of these systems performed worse, post release, than the original system.
Another classic example is an intermarket trading system for Treasury bonds. The intermarket divergence concept is based on the 30-year T-bonds and the Philadelphia Utility average, as follows:
- If Bonds < Average(TBond,6) and UTY > Average(UTY,20) then buy at open
- If Bonds > Average(TBond,6) and UTY < Average(UTY,20) then sell at open
This system is simple but shockingly effective. From Sept. 22, 1987, though Aug. 7, 2014, including $50 per trade for slippage and commission, the system made $250,768.75 trading a one-lot with 60.3% winning trades and an average winning trade of $1,967.30. The maximum intraday drawdown was $22,131.00. The drawdown is a bit high, but the system has remained robust, working well since originally published in 1998.
The original parameters developed in 1998 were eight and 18. The above results are based on six and 20, as stated in the rules. If we add a “disaster” stop at $5,000, it has a minimal effect on our profit and drawdown, but drops the largest losing trade from –$16,000 to only –$5,081. We still make almost $240,000 with a drawdown of a bit over $24,000. This is slightly worse performance overall but is necessary to make the system more tradable. The largest losing trade has a significant effect on the total risk-adjusted returns of a trading system.
It’s tempting to look for additional filters, such as a correlation filter, a volatility filter and so on. Doing this is tempting but may not have the desired results. It may reduce drawdowns on the backtested results but those improvements will not necessarily continue into the future.
Complex filters may look better during the sample period—the system becomes designed for the precise history on which it's developed. Further, filtered results often are not statistically significant. A filter could take out 30 trades, with only two to three being large losers. To determine a filter’s viability, it must pass statistical tests showing that the population of the filtered trades is worse than the population of trades within the system. If we can’t say conclusively that the filtered trades are worse, we shouldn’t trust the filter.
While the simpler systems might be useful to a fund trading large amounts of money, they might not work well for those operating with less funds. For those trading between $20,000 and $50,000, they might focus on drawdown or the largest losing trade. This is incorrect. It’s unlikely these statistics, which could be major outliers, will repeat immediately. For the T-bond system, it might be reasonably assumed that the drawdown would demand a $50,000 account. However, the drawdown could be directly traced to the 2008-09 financial meltdown—a once a generation event. Without that outlier, a $25,000 account is more reasonable.
One concept that provides a more realistic view of risk is start trade drawdown. This was developed by Keith Fitschen and is available in TradersStudio. The idea is as follows: For each day during the backtest, we ask the question, “If I started trading on this day, what is my maximum out-of-pocket during the rest of the period?” Let’s take a look at this statistic for our first iteration of the system without the $5,000 disaster stop (see “Start trade drawdown,” above). The same chart, but with the $5,000 disaster stop in place, is shown in “Limiting losses” (below).
Risk of ruin
Risk of ruin is a concept that considers both account size and system performance. In general, a system with a high drawdown thst is more likely to perform as expected would produce a lower risk of ruin. This is the point at which our starting account size is such that we are no longer able to trade. For example, we can define “ruin” as losing 50% of the account value, and “risk of ruin” as the probability we will do that. Risk of ruin drops as we gain capital and is important for smaller accounts.
Perry Kaufman, in “New Trading Systems and Methods,” presents a formula to calculate risk of ruin. This formula is from Ralph Vince’s “Portfolio Management” and is a summary of P. Griffins’ work in the theory of blackjack gamblers press published in 1981. It is a fair approximation of risk.
This formula is as follows:
This method gives us a good estimate of risk of ruin. For the purposes of this example, let’s assume we have a $35,000 account that is under capitalized for a system with a $22,000 to $24,000 drawdown. Consider our T-bond system without stops. Our largest losing trade is $16,380. The win/loss ratio is 1.21, and our winning percentage is 60.32%. Let’s define a 70% drawdown as ruin. In this case, our risk of ruin is 44.8%, so more than 55% of the time, we could trade this system with a $35,000 account.
Now consider the system with the $5,000 stop. Our win-loss ratio drops to 1.18 and the winning percentage drops to 59.4%, but our largest losing trade is only $5,081. This has a huge impact on our risk of ruin, dropping it to only 9.42%. More than nine times out of 10, we would be able to trade it with a $35,000 account. With even a $25,000 account, the percentage is only 18.45%, so four out of five times we can start this system that has averaged $8,800 per year with a $25,000 account.
You must be careful not to reduce the theoretical risk of ruin too much, however. Doing so may reduce the reliability of the system in the long term. As always, we should not just focus on one metric, but look at the system as a whole. This calculation needs to recognize that trading is a sequence and that if we make enough money early, we alter the risk of ruin.
We may reduce our risk of ruin calculation if we assume that there is a serial correlation in trades. For many trend-following systems, we could wait until the system is in drawdown to start trading. Consider a simple example of a channel breakout system on the yen.
This system only enters a trade when the current theoretical equity curve is below the average curve. Simply put, after a period of losing trades, we start trading. We will test this system on the yen from Jan. 1, 1990, to the present day with $50 slippage and commission.
We use a 30-bar breakout for the core system. The results of this core system aren’t very good. The system made $45,500 trading a one-lot with 33.9% winning trades and an average winning trade of $6,864.06 and an average losing trade of $2,937. The maximum intraday drawdown was $65,088.
Here are the results of trading when the current equity curve is below its moving average. We will only trade when the current equity is below a 130-day moving average. Our results are significantly better due to the serial correlation in trade performance. The net profit was $127,587.50 with a 42.3% winning percentage. The average winning trade was $8,001.67 and the average losing trade was $2,743. The maximum intraday drawdown was $30,887.50.
Because the win-loss ratio and the winning percentages are much higher, the risk of ruin is much less. The largest losing trade is even smaller. The unfiltered system assumes a $35,000 account and has a risk of ruin at 67.6%, which isn’t good for these parameters. With the filter, the risk of ruin drops to 24.4%, which is manageable. This is a better set-up for a $35,000 account.
The problem is that this concept only works when we have serial correlation in trades. If we don’t then we can’t use any type of performance feedback to make systems more tradable.
Sizing with the account
Finding a way of allowing smaller accounts to trade more robust systems is an important area of research. Often, we don’t have serial correlation in trades so we can’t use simple equity curve feedback.
Ralph Vince has developed math that helps define risk. He calls it “leverage space.” This allows us to answer the question “what are the odds this system will be profitable after 12 months?”
Vince is most famous for the money-management methodology known as optimal f, which has some limitations. First, it cannot work on a portfolio and, second, it is too aggressive. Vince recognized these limitations and developed leverage space as an extension. One concept he developed is called the inflection point. This is the point of diminishing returns whereby optimal f risk is increasing faster than the increase of returns.
There are two inflection points: One on the left side of the optimal peak and the other on the right. We are interested in the less aggressive left-hand side of the peak. Indeed, the inflection point offers the best risk-adjusted value. Because we are to the left of optimal f, if the system slips, it will be, at worst, at the peak or slightly to the right. The value is based on the recent sequence of trades, so it can be calculated in real time and can adjust to changes in trade distributions and correlations.
We can use this inflection point, as well as optimal f, to turn a system on and off so that it can be traded with a smaller account. (Vince contributed to this idea when contacted for this article.) His idea was to define two windows. First, we have a long-term time window for length A and a shorter-term one for B. We then keep track of optimal f for both windows, A and B, as well as the left hand inflection point for B. When the optimal f for B >= A, we stop trading. Conversely, we start trading when optimal f for B <= inflection point of A. This works in a more general case than equity curve feedback because it does not depend on trade correlations.
The business has changed. Many traders with larger accounts are gone, replaced by smaller accounts. These smaller accounts make it harder to trade robust systems. Developing new concepts in money management is key in allowing us to define risk of ruin for any size account .
Systems that are fully disclosed and understandable to clients are best. Systems vendors need to look out for the client and understand risk. First and foremost, plans must be designed for systems that have defined levels of risk based on trader resources.
Murray A. Ruggiero Jr. is the author of “Cybernetic Trading Strategies” (Wiley). E-mail him at firstname.lastname@example.org.