I’m sure we all want to create trading strategies that perform better and last for longer but there are a number of issues we need to look out for when developing robust trading strategies, some are well-known and some perhaps aren’t.
In this episode we’ll be talking with Perry Kaufman about strategy development and more specifically some of the issues that can catch us out when creating trading strategies. Perry raises some interesting points about optimization that may not be well known plus he shares loads of tips to creating more robust strategies.
Perry writes extensively on markets and strategies, having published fourteen books and has just released a new book on building algorithmic trading strategies, which we’ll be discussing in this episode.
He has worked and consulted to a number of successful CTA, investment and prop trading groups, creating systematic trading and hedging programs.
This is also his 2nd appearance on the podcast, appearing as a guest way back in Episode 10.
Sponsor
Topics discussed
- The most robust type of systems
- How your choice of optimization values could be misrepresenting your results and how to choose parameters that give a more accurate picture
- The mistakes traders make when analyzing optimization runs and tips to doing it properly
- How to really determine if a new trading rule is robust
- Reducing risk by using multiple parameters
- What the number of profitable runs in an optimization can tell you about the robustness of a strategy
- Why diversifying across strategies instead of across markets could be a better approach
- The challenges of building robust strategies using Genetic Algorithms and Neural Networks
Resources mentioned in this episode
- To learn more about Perry and his work, checkout his website Kaufmansignals.com or Perrykaufman.com
- He can also be contacted on perry@perrykaufman.com
- Books mentioned in the show:
 |
 |
Quotes
Top tips
Perry shared a number of great tips, here are some of my favorites:
- Use average of results in an optimization run
- Remove strategy rules that don’t add much value
- Look at the % of profitable runs in a distribution, Perry personally likes to see 66% or higher runs to be profitable
- When considering the values in an optimization range, look for the values to be evenly spaced %-wise, for example, 10/20/40/80, not linear like 10/20/30/40/50 because linear can make the larger values look more stable purely because they’re closer together as a %
- When adjusting strategy rules, don’t be fooled by a few results going up, you want something that is generalized and improves most cases
- Neural networks/genetic algorithms – feed in fewer items to get a more robust answer
Transcript
Got A Question, Topic or Guest you want to see on the Podcast?
Do you have a specific question, topic or guest you’d like to see on a future podcast episode?
Click here and submit your suggestion for a chance to have it featured on an upcoming podcast episode.
Subscribe to Better System Trader and never miss another episode!
 |
 |
 |
 |
Please support the podcast by giving an honest Rating/Review for the show on iTunes!
Episode Released:
3 April 2016
006 – Dr Howard Bandy
058 – Adapting to market conditions with John Ehlers
141 – Models are like unicorns with Kevin Saunders
079 – Strategy Validation with Dave Bergstrom
010 – Perry Kaufman
Great Interview as always , Isn’t it advisable to first wet one’s hand with discretionary trading to get a feel for momentum & mean reversion in a simulator or 1 mini lots in real time environments before heading onto rule based discretionary & finally into systematic trading ?
Hi Django,
There is a benefit in actually trading, especially if it’s small volume, but discretionary trading does not (in my opinion) give you any leg up for systematic trading.
I would suggest you consider a simple, long-term moving average system on a market that has a history of success (either long-term bonds or EURUSD). That way you can learn to follow the system, not always an easy task, and assess whether it is performing as expected. Systematic trading is a very different skill set and discipline than discretionary trading.
And, of course, start as small as possible so you’re not worried by any losses.
Perry.
What’s the difference between long term data points and short term data points(say comparing one sec to one day)? I still dont understand why long term strategy dont work on shorter timeframe
That’s a good question and goes to the heart of why we use high-frequency (intraday) or low-frequency (daily or weekly) data for different systems. The closer you look at the data (e.g., higher frequency) the more “noise” you see, that is, more changes from up to down. If you put up 5 min data on a screen it will look very noisy. Changing that to 30 min data smooths out the appearance. Change to daily and weekly makes it even smoother.
So the process of skipping data is that same as smoothing the data. If you try a 52-week moving average system on (of course) weekly data, then try a 252-day MA on daily data and a 2016-hour MA on hourly data (assuming 8 trading hours) you would expect to get the same results, but you don’t. The amount of noise in higher-frequency data will cause the trend to change much more often, so the results will deteriorate as you look shorter data intervals.
Then there is the separate issue of whether there are trends in short calculation periods. We can explain the trends in longer time periods by the influence of the Central banks. They manage the interest rates, which then affect FX, and carrying charges of commodities, and finally the equity markets (borrowing, etc.). So longer-term trends have worked forever. But what about over 10 or 20 days? There is no “policy” that keeps repeating over that time frame. Short-term moves are driven by news, earnings, supply-demand disruptions, and other short-term events, many of which affect the market one day, then are discounted in a few days because the market has a bad memory. Those moves, looking back, seem more like noise.
My own work concludes that a trend system is best applied to longer calculation periods, and mean-reversion to short periods (to take advantage of the noise). Normally, the equity index markets are the best candidates for noise, which the short-term rates are the best for trends. So it’s a philosophy that you might develop over time. That doesn’t mean that you shouldn’t try to find a better solution, or a way around the problem. It’s just nice to have a starting point.
Best regards,
Perry