Optimal Stop-Loss and Take-Profit Levels

[dacoffey]

I am working on constructing a trading system using neural networks to generate signals for intra-week trades.
Results have been promising so far.

The system uses both trailing stop-losses (short and long) and take-profit (after a certain PL is achieved for a trade).

Currently, I am trying to tweak the levels to give optimal PL.

Using the past month of data, I tested my system with various stop-loss and take-profit levels.

My results seem to indicate that for this system, optimal stop-loss level is around 35 pips and optimal take-profit is after 50 pips.

Marginally higher PL can be achieved by greatly increasing the stop-loss, but my heart would prefer tighter stops and slightly lower profits.

How do my results compare to others?
I'd like to see what levels people utilize for their systems.

[dacoffey]

Total PL vs Stop Loss

[dacoffey]

Oh yes, forgot to mention:

In the graph where Stop-Loss is being changed, the Take-Profit was held at 50.

And in the graph where Take-Profit is being changed, Stop-Loss was held at 35.

[bubo]

very interesting, thank you for sharing.

i find that limits and stoplosses greatly vary among time scales and currency pairs involved. it also matters what strategy you use - i.e. at what point you enter and quit a trend, and how your stops and limits will change that pattern.

what currencies, time scale, and strategy have you been using to test this?

[dacoffey]

I also forgot to mention that the take-profit and stop-loss calculated for each pair takes the spread into account.
I figured this would provide more consistent results between pairs.

My nets are trained to find moves up or down within the next 4 hours, and my tick data is 10 seconds.
An up, down, or steady signal is generated for 18 pairs.
Each pair has a signal threshold that must be tuned.
So far, I can only get profitable results from 8 pairs, so actual trades are limited to these pairs:
EUR/USD, USD/CHF, AUD/USD, USD/CAD, NZD/USD, EUR/GBP, EUR/JPY, EUR/AUD, and AUD/JPY

Usually my trades last around 2-16 hours before one of 3 things happen:
* Stop loss gets hit
* Take profit gets hit
* Nets give a reverse signal

The backtest data I presented only covers the past 3 weeks, and is not continuous; unfortunately there are some gaps.
The PL generated comes from 25 different trades (16 long, 9 short) over this time period that were signalled by my neural nets.

I am still working on collecting data and hope to post more analysis later.

[bubo]

so your charts are on 4 hour candles?

how do your nets decide to signal long or short? did yo describe this somewhere?

sounds very interesting, and i like the fact you are proceeding so systematically.

[dacoffey]

I don't really use charts.
I find when I look at them, I start second guessing my system.

The nets have been trained using past data to find big moves - anything > 3 * pip spread.

The inputs to the nets are simple indicators such as MACD and RSI.

[anomaly]

dacoffey--
I'm very interested in this topic as well. I have backtested many systems with exit strategies similar to what you're testing on your system right now.

I would agree with bubo in that I was not able to find one optimal stoploss/takeprofit level for all currency pairs. In reality, they varied greatly. With one system, I was able to get stops as tight as 15 pips on the EUR/JPY (great entry timing), but could never get anything closer than 30 on the GBP/USD. Some crosses simply have different noise levels.

The conclusion that I came to after testing these systems extensively (over 8 months of historical data) is that I was trying to overoptimize my exits. Unless you have a strategy that provides you with specific exit points (like Elliott Wave or retracement systems), it's very difficult to project the strength of a move. When you overoptimize your exits, you're trying to do just that--determine the most frequent (optimal) levels of market strength.

I think this is a good discussion, though, and I'm happy to find someone else that is systematically testing his theories in the same sort of ways.