1) The document discusses methods for evaluating the performance of trading systems and decomposing their returns into different components including skill, trend, and bias.
2) It presents equations for estimating the bias from training on randomly permuted data, and adjusting total returns to remove the estimated bias and isolate the skill and trend components.
3) Applying these methods allows traders to judge a system's true skill at finding patterns in the market independently of any gains from taking advantage of long-term trends or biases from overfitting to random patterns in the training data.
1) The document discusses methods for evaluating the performance of trading systems and decomposing their returns into different components including skill, trend, and bias.
2) It presents equations for estimating the bias from training on randomly permuted data, and adjusting total returns to remove the estimated bias and isolate the skill and trend components.
3) Applying these methods allows traders to judge a system's true skill at finding patterns in the market independently of any gains from taking advantage of long-term trends or biases from overfitting to random patterns in the training data.
1) The document discusses methods for evaluating the performance of trading systems and decomposing their returns into different components including skill, trend, and bias.
2) It presents equations for estimating the bias from training on randomly permuted data, and adjusting total returns to remove the estimated bias and isolate the skill and trend components.
3) Applying these methods allows traders to judge a system's true skill at finding patterns in the market independently of any gains from taking advantage of long-term trends or biases from overfitting to random patterns in the training data.
markets is the Skill component, which cannot exist in permuted data. In other words, Equation 10 below represents the average total return obtained from training on permuted markets. As a point of interest, look back at Figure 16 here. That figure is based on profit factors, while the equations just shown are more applicable to total return. Nonetheless, the principle is the same for either measure of performance. The bars in this histogram represent the values of Equation 10 above (plus the single observation for the unpermuted market, which is also used to compute the histogram). The red vertical bar is the value represented by Equation 9 above. We can thereby see the effect of Skill on performance relative to the performance of the trained trading systems when Skill is not involved. If the Skill component is significant, we would expect the vertical red bar to be on the far right side of the histogram, pushed there by the ability of the trading system to find authentic patterns in the market. Conversely, if the Skill component of performance is small, the red bar is more likely to be somewhere in the interior mass of the histogram. Additional information can be gleaned from the permutation training by analyzing Equation 10 in more detail. When the trading system is trained by maximizing performance on a permuted dataset, the training algorithm will attempt to simultaneously maximize both components of Equation 10: it will try to produce an unequal number of long and short trades (assuming that the user allows unbalanced systems) in order to take advantage of any long- term trend in the market(s), and it will try to take individual trades in such a way as to capitalize on any patterns it finds in the data. Of course, since it is operating on randomly permuted data, any patterns it finds are inauthentic, which is why the Skill component plays no role. Now suppose we have a system that makes only long trades, and it does so randomly; individual trades are made with no intelligence whatsoever. Equation 11 shows its expected return: In this equation, BarsLong is the number of bars in which a long position is signaled, and TotalNumberOfBars is the number of bars in the entire dataset whose performance is being evaluated. This ratio is the fraction of the number of bars in which a long position is signaled. Looked at another way, this ratio is the probability that any given bar will signal a long position. TotalTargets is the sum of the target variable across all bars. This will be positive for markets with a long-term upward trend, negative for those with downward trend, and zero for a net flat market. So if our system is without intelligence, with signals randomly given, we will on average obtain this fraction (the fraction of time a long signal is given) of the total return possible. A similar argument can be made for a short-only system, although the sign would be flipped because for short systems a positive target implies a loss, and a negative target implies a win. Combining these two situations into a single quantity gives Equation 12, the expected net return (the Trend performance component) from a random system that contains both long and short signals. When we train a trading system using a permuted market, we can note how many long and short trades it signaled and then use Equation 12 to determine how much of its total return is due to position imbalance interacting with long-term trend. Anything above and beyond this quantity is training bias due to the system learning patterns that happen to exist in the randomly shuffled market. This training bias can be estimated with Equation 13. It would be risky to base a Bias estimate on a single training session. However, as long as we are repeating the permutation and training many times to compute the probability of the observed return happening by just good luck, we might as well average Equation 13 across all of those replications. If at least 100 or so replications are used, we should get a fairly decent estimate of the degree to which the training process is able to exploit inauthentic patterns and thereby produce inflated performance results. Two thoughts are worth consideration at this point: • TSSB allows various BALANCED criteria which force an equal number of long and short positions in multiple markets. In this case, Trend will be zero. • Powerful models will generally produce a large Bias, while weak models (not necessarily a bad thing) will produce small Bias. Now that the Bias can be estimated, we can go one or two steps further. Equation 9 can be rearranged as shown in Equation 14 below. This equation shows that if we subtract the Bias estimated with Equation 13 from the total return of the system that was trained on the original, unpermuted data, we are left with the Skill plus Trend components of the total return. Many developers will wish to stop here. Their thought is that if a market has a longterm trend, any effective trading system should take advantage of this trend by favoring long or short positions accordingly. However, another school of thought says that since deliberately unbalancing positions to take advantage of trend does not involve actual trade-by-trade intelligent picking, the trend component should be removed from the total return. This can be effected by applying Equation 12 to the original trading system and subtracting the Trend from the UnbiasedReturn, as shown in Equation 15 below. This difference, the Skill component of Equation 9, is often called the Benchmarked return because the NetExpectedReturn (Trend component) defined by Equation 12 can be thought of as a benchmark against which to judge actual trading performance. These figures are all reported in the audit log file, with p-values for the profit factor and total return printed first. Here is a sample: Net profit factor p = 0.0600 return p = 0.0400 Training bias = 52.3346 (67.1255 permuted return minus 14.7909 permuted benchmark) Unbiased return = 55.4320 (107.7665 original return minus 52.3346 training bias = skill + trend) Benchmarked return = 39.8372 (55.4320 unbiased return minus 15.5947 original benchmark = skill) The Training bias line is Equation 13, with the term benchmark referring to Trend in the equation, and these figures averaged over all replications. This figure (52.3346 here) is the approximate degree to which the training process unjustly elevates the system’s total return due to learning of inauthentic patterns. The Unbiased return line is Equation 14, the actual return that we could expect in the future after accounting for the training bias. This figure includes both the true skill of the trading system as well as the component due to unbalanced trades that take advantage of market trend. The Benchmarked return line is Equation 15. This is the pure Skill component of the total return.