Optimisation of a system issue

I am currently testing and optimising a trading system over 1 year of data, and sorting the results by Gross Profit Made.

What I am finding is that by sorting the results by Gross Profit Made, the system has long periods of small losses then one big gain. Although over a year this provides a good return, drawdowns in the interim are bad - and I am looking for regular cashflow with lower drawdowns rather than the largest gain made over a year.

Can anyone think of a way to optimise results for maximal cash-flow each month rather than Gross Profit Made in a year? Is there a mathematical formula I can use?

I tried using a average of x bars, but this still doesn't solve the problem, eg:

Week 1: -$40
Week 2: -$40
Week 3: $8000

whereas I would like more:

Week 1: $900
Week 2: $1500
Week 3: $2000

(this is a very simplified example but illustrates what I am after).

Many thanks,
Alex

ive faced similar problems in the past; i found that the best way to solve this issue was to code the money management directly into the optimization routine, and then sort by max returns. using a % based MM routine works best, because large drawdowns will inhibit the account from growing very large, whereas strategies with consistent profits and low maxDD will fair the best.

Without knowing how your platform calculates Optimized Backtest Results, The problem OPTIMIZING with most backtest engines is that they just crunch numbers with no regard for anomalies.

Here's what I mean. eg. BMC/USD {fake pair}
News Release: Date Nov 15 2006.
The Country of BMC has just discovered Crude Oil rising through the bedrock like a meadow freshwater spring.

BMC/USD Spikes up Big Time. Huge Trade.

News Release: Date Nov 16 2006.
The Country of BMC has just discovered that the Crude Oil rising through the bedrock like a meadow freshwater spring is the result of a broken pipeline in the neighbouring country of FKU.

Guess what happened to BMC/USD.

So here's the Problem with Optimizing.

Any set of numbers you achieve will very probably include a Buy before or shortly after BMC found the Oil But that's not all... they will also likely include a Sell before or shortly after BMC discovered the Oil was not theirs.
Do YOU think this might skew your numbers?

Hi,

Thanks for your replies - I will code money management in directly.

In response to Bemac's comments, my system is taking advantage of the volatility at the start of the US session (GBP:USD) at 13:00 GMT. However, agreed - skews in data will occur.

Has anyone successfully managed to use an optimised automated system?

How about, for example, optimising for %wins and risk:reward rather than gross profit - would this change things?

Alex

Hi Dac_uk

Some possible ideas:

Higher number of transactions relative to time can have the effect of smoothing cash flow (e.g. if trading 5 times per day, 5 consecutive losses means one negative day; if trading 5 times per month, it means negative income for a whole month). However, in trading a shorter time period, we need to consider both the increased volatility (e.g. news announcements) and the higher relative costs (i.e. 4 pip spread is more significant over a 20 pip move than a 100 pip move).

In similar vein, consider trading simultaneous uncorrelated positions to improve diversification (overlapping wins and losses).

All trading systems are ultimately a compromise between win rate and win size. What you appear to be saying is that you have a low win rate (several consecutive losses) but high win size (8,000 vs –40 in your example). Scaling out using tighter profit targets (if you use them), wider protective stoplosses, and/or moving stoploss more quickly to breakeven can all improve win rate, but at the cost of win size. (As an extreme example, a system that takes profit at 20 pips, but uses a 200 pip stoploss, will have a very high win rate). However, much of all this depends on whether we’re trend-following, trading breakouts or attempting to fade extremes.

If losses are clustering, is there evidence to suggest that your system fares poorly during certain market conditions (e.g. trending, ranging, …), and is it possible to detect when such conditions occur, and downsize positions (or not trade at all)? Also, how selective is your system in choosing the highest probability setups (could some of the losers be eliminated)?

Good luck
David

Alex, you may be faced with a relatively unsolvable problem (at least on a realistic basis) because most ccy pairs seem to trade smoothly for a while, interspersed with significant (and often unpredictable) jumps. $CAD is a somewhat clear example - if you look over the charts over the last few months you'll see lots of flat periods followed by a sudden drop, so if you are trading the pair, you'll have areas of small - but consistent negative returns, followed by a significant win. (Assuming that you are shorting the position) How you get around that is really a matter of how big your purse is - for example, by trying to minimize the correlated ccy pairs in your portfolio as possible, scaling in and out of trades, including options, etc.

Thanks - this is all food for thought and helpful.

You should be able to optimize on fitness functions other than profit. This will "steer" your results in a particular direction. For example, you might get a better curve with a fitness function of netprofit/drawdown or netprofit/trade or expectancy, or netprofit/margin.

In my humble opinion programming in money management in your initial optimizations is a big mistake. You should develop the best possible equity curve using a fixed position size. If you have a positive expectancy, money management will only make it better. If you have an ugly curve, money management will cover up some of the areas which would otherwise be identified and improved upon.

I know it takes time but I've spent some time on this over the years.
Simple Fictional Example of Spike Influence on Optimisation.

Would you rather be Trading Optimized Set # 11 or Set # 18?