Trading System Development

http://www.blueowlpress.com/WordPress/feed/

Saturday, 19 of May of 2012

My book Quantitative Trading Systems (QTS) describes in detail the trading system development process I recommend.
Modeling Trading System Performance (MTSP) is a sequel to QTS and extends it. MTSP begins with the results of a trading system and analyzes the performance.

The outline presented here gives a framework for blog postings.

From “Quantitative Trading Systems”

1. Define the objective function.

2. Decide what to trade and how to trade it.

3. Design the trading system.

4. Determine the length of the in-sample period.

5. Determine the length of the out-of-sample period.

6. Decide what to optimize.

7. Perform walk forward runs.

8. Evaluate the out-of-sample results.

9. Trade the system.

10. Monitor the results.

From “Modeling Trading System Performance”

11. Treat trading as a business.

12. Understand the distribution of the trades.

13. Determine the system’s risk.

14. Consider the trader’s tolerance to risk.

15. Perform Monte Carlo simulation.

16. Determine position size.

17. Evaluate system health.

You need to be loged to make a comment

Comments RSS TrackBack 3 comments

Log in to Reply

charlie

in November 17th, 2011 @ 14:55

I am looking forward to the discussion about walk forward analysis.
Log in to Reply

heuy

in May 7th, 2012 @ 06:55

Dear Dr. Howard:

I have questions on Monte Carlo.
1. Can Monte Carlo be used as a tool for system validation?
2. If yes, how MC can go together with t-test, chi-square and control chart in system validation?

Log in to Reply

HowardBandy

in May 7th, 2012 @ 08:09

Yes, Monte Carlo can be used as a tool for system validation. As a first step, I recommend establishing the “best estimate” set of trades through walk forward testing, including live trades when they are available. A quick visual inspection of the out-of-sample equity curve will let you know whether to proceed further.

Statistical tests, such as the t-test, and control charts, as used in manufacturing processes, depend on the standard deviation being small relative to the mean. One of the difficulties encountered when working with trading systems is that the standard deviation is typically the same magnitude or larger than the mean. For example, a trading system might have a mean of 0.1% per trade with a standard deviation of 0.2% per trade. Compare with manufacturing of ball bearings, where the mean is 0.25 inches and the standard deviation is 0.001 inches.

As I describe in Chapter 11, “Is it Broken?” of the “Modeling” book, pages 221 through 272, traditional statistical tests and control charts can be constructed using trading results. Using the control chart as an example, the control band is formed around the desired value by adding and subtracting some number of standard deviations. The difficulty when using this method on trading results is that the control bands are so wide that they are not very useful. The bands usually include 0.0, meaning that a system could be showing random or negative results and still be within control parameters.

The t-statistic is computed as the ratio of mean to standard deviation, multiplied by the square root of the number of data points used. It is still the case that the standard deviation is greater than the mean. But including the number of data points provides a way to determine whether (that is, determine the probability with which) the mean is greater than 0.0. The length of the window over which the statistic is computed causes a lag, similar to the lag in moving averages. When applied to a time series this lag causes the statistic to be measuring the entire group of data, centered half the number of data points in the past. If current results are deteriorating, there will be a lag before the statistic confirms that. As a rule of thumb, a t-statistic greater than 2.0 gives confidence that the mean is greater than 0.0, meaning the trading result is better than random.

Since trading systems are dynamic, with results rising and falling with both the opportunity for profit available in the data and the synchronization between the model and the data, methods used to measure system health should be dynamic as well. Monte Carlo analysis gives the opportunity to simulate profit and drawdown using the recent trade results. As the system performance changes, the most recent trades are incorporated into the Monte Carlo analysis and can help measure system health and determine position size. Contrast this with static analysis where the system is assumed to always be in the same state of health and position size is determined without regard for changing conditions and changing performance.

I hope this helps,
Howard

Quantitative Trading Systems

Login

Recent Posts

Recent Comments

Blogs worth exploring