Include ATM parameters in optimization / walk forward analysis?
Hi all, I have a question about whether to include the ATM settings as part of the strategy optimization and walk forward analysis.
We're currently working on a strategy that implements BigMike's excellent ATM script with 3 profit targets & progressive stop-loss settings. My question is whether to include the PT's & SL parameters as part of the optimization and walk forward analysis.
The strategy has 4 parameters with static ATM, 8 if PT''s & SL are included. The concern is that we'll overfit if we use too many parameters, however, keeping PT's & SL is important in adapting to different market conditions.
This post has been selected as an answer to the original posters question
Why don't you backtest the three single components of the strategy first?
If you evaluate a strategy with 3 different profit targets, this is similar to running three simple strategies in parallel. The three of them enter the market at the same time but exit under different conditions.
Before evaluating the complex strategy which enters and exits three different positions, I would definitely evaluate each of the component strategies in order to find out whether they are profitable and robust.
If one of those is not performing well alone, I would eventually not use it for the basket.
The potential benefits from blending different strategies into one would arise from a reduction of the variance of returns. But before focusing on the variance, I would like to see the expectancy and key figures of the component strategies.
Include Profit and Stop Loss Parameters With the Optmization
I would definitely run an optimizer over different values for the profti target, stop loss and trailing stop, but not before I have tested the system for stability and added filters.
By the way there is a little book, which I would like to recommend, as it is well structured and an easy read. I have no relation whatsoever with the authors, but it is the best introductory book for developping trading systems that I have seen.
Urban Jaeckle + Emilio Tomasini: Trading Systems
6 Steps to Develop a Trading System
The authors recommend to follow the steps listed as below for developing a trading system (see page 87):
1. Ideas. Code and define entries and exits.
2. Add slippage and commissions.
3. Stability Tests / Check for stable area or input parameters.
4. Add filters such as time filters, trend filters and volatility filters.
5. Analysis of MAE (maximum adverse excursion) and MFE (maximum favourable excursion).
6. Optimization of stop loss, trailing stop and profit targets.
I think it is a good idea to follow this approach.
It seems that you are starting with the last step....
The following 2 users say Thank You to Fat Tails for this post:
Hi Fat Tails and many thanks for your comments and book recommendation. I'll check it out, so far I've been through Robert Pardo's book on Evaluation and Optimization of Trading Strategies.
My strategy concept is to use market sentiment as basis for a trend following model. To do this I use the SentiTrade indicator which measures the real time ratio between positive and negative news stories. BTW, in the interest of full disclosure, I'm one of the founders of SentiTrade :-)
The assumption is that market psychology at certain times is a trend driver. SentiTrade scans news stories related to equities markets and I'm therefore developing this for ES, YM, NQ and TF. These following are my entry conditions:
- If the ratio of pos vs. negative financial news stories is above average
- And there's positive sentiment momentum
- And the price close above average
- Then go long
- If the ratio of pos vs. negative financial news stories is below average
- And there's negative sentiment momentum
- And the price close below average
- Then go short
Next I discovered that the conditions worked best at certain times of the day and added a filter for that.
I then tested for robustness with NinjaTrader's Walk Forward Analysis; 5 sentiment average settings, 5 degrees of pos/neg momentum, 5 time periods in which momentum was calculated and finally, 5 time periods in which the price average was calculated.
A large historical data set was used (3 years using 1 minute bars) to make sure I had sufficient "degrees of freedom" and that all market types were represented. I also made sure that a statistically sound sample of trades were generated, and that slippage and commission costs were included.
This all came out with acceptable results using a static PT's & SL with a 3:1 Risk Reward Ratio. Including the PT's and & SL parameters in the walk forward analysis however, does not improve performance, to the contrary.
I optimized PT's & SL in steps of 5, to keep all strategy parameters proportional. The "genetic" optimization option was used and I now wonder whether it has something to do with it, i.e. if some of these settings (crossover rate, generation size etc.) have to be modified when adding parameters?