Anyone used ARMA + GARCH models in forex trading?

Hi All,

I am currently exploring using ARMA + GARCH to predict next days position.

Right now I'm running some models on the last two weeks. It will take some time to complete so I thought I would ask the forex factory community if anyone has explored these yet?

Looks to have provided some good results on the SPY in a post I have read on http://www.quintuitive.com/2012/08/2...s-for-trading/ and was wondering if anyone was able to get the same result with currency pairs?

I'll try and post my results when they are done computing.

For the two week test it looks great!

The test tried to predict the next days return and returned either a 1, 0, or -1 then based on that value would open a trade long short or no trade.

The arma+garch model takes looks at the past 40 bars and choses the arma model that has the lowest AIC. Then apply a garch 1,1 to account for volatility.

The blue line represents the closing price returns of the EURUSD 4 hr chart as if you were to purchase and hold a long trade.
The green line represents ARMA+GARCH Model in which a long or short trade was purchased.

This is only around two weeks 72 bars worth of data being analyzed looks promising.

The ARMA model alone on quite a bit of data doesn't look so good.

Right now running the same data here through with the GARCH model for volatility. Could take 24hrs + to finish so will be awhile before next post and I can see the results

That's way too small a window for GARCH. It needs something like 2000 bars to converge.

Yeah I was just reading that most of the time anything with less then 500 points will give rubbish.

I have already kicked off the the building of the forecast models. Only a few hours in. Wonder if I should stop and rerun with a larger window or wait until it finishes and see what the output looks like lol.

Probably you should because I have another bad news for you! Because of the fractal nature of the financial time series, you often have a fractional integration order (google for it). That is the d in ARIMA(p,d,q) is not an integer. Of course an I(d) model cannot handle it. The good news is that the AR(p) part acts like a differentiator and the MA(q) part acts like an integrator. The model will compensate for the over/under differentition with either more p or more q. Of course you will need much more samples to fit the model...

I've never been able to figure out a way to use ARIMA with FX. Even predicting the next bar with very high acuracy doesn't make the job done. I read a paper where their could forecast the next bar with 80% acuracy (neural net, not arima) yet their could never build a profitable system out of it.

I restarted it. I am using history of a 1000 point now. However, Hasn't completed yet lol the forecast results over a period of time.
Will post them If I don't lose power have to start over again lol. Hopefully will give some better results with more data points.

Strange about the 80% accuracy I think you would be able to create a profitable system on that.
If someone couldn't build a profitable system knowing which way the market would go with 80% certainty that doesn't leave much hope for the rest of us .

Happen to have a link to that paper? Would be an interesting read if you could find it.
To see the road block they ran into.

I didn't keep the paper because it was about Neural Nets that I consider just like curve overfitters. They didn't explain how they tried to use their forecaster.

I think the failure comes from the fact that even if you know for sure the next bar will be green, it doesn't prevent the two subsequent bars to be red. The next bar can be up in a strong sell-off. But you can't change your mind every bar you need to let the trade grow. The method was local and a local method will always fail because the entry will always be the same once definced: either a breakout or a fading strategy. With or against the move. There are only two choices. The first fails in ranges the other fails in trending markets. Market alternates the two. Both fail. Global approach (Big Picture) is required IMO.

In a 1000 bar period the market will certainly have already changed. The financial time series aren't at all stationary even after de-trending/diff'ing because of the heteroscedasticity. That's why I'm afraid ARIMA isn't very useful.

Why is your fitting so slow? Did you try online estimation with kalman filters (see the attached paper)?

Not sure why taking so long this go around maybe since i increased the history so much. It's forecasting every bar and then refit's every bar.
Runs through a series of arima models (0,0) -> (5,5) so each bar generating something like 35 models and picks the best fit out of those based on the AIC. Then fits a garch (0,0) to just make sure volatility was in there and to keep from having to do additional computation I don't cycle through orders for garch.

Fairly heavy computational wise generating the backtest. Even running in Linux to utilize parallel processing....

Been running for a few days now so hoping it will complete soon and I can post the results taking a longg time though.

I think GARCH(0,0) doesn't exists... GARCH(1,1) is good enough for financial time series. You can just postulate this.

Why don't you use ACF and PACF to guess the p,q? => http://people.duke.edu/~rnau/411arim3.htm
I'm not sure but p and q aren't supposed to change often, do they?

If you really want to estimate the coefficients online (bar-by-bar), you should really consider a Kalman filter.

Yes, GARCH(1,1) is promising in Finance and Econometrics.
you dont need to waste time with the ACF and PACF, Eviews computes that accurately, to the best of my knowledge

I think the problems with trading can be greatly reduced if we can clearly differentiate TRENDING and RANGING periods.
I have started using the GARCH(1,1) to estimate volatility.
Then I plan using Neural Networks (Pattern Recognition) to forecast.
But I think it will be beneficial to have something like a GARCH Oscillator.

Is their any GARCH oscillator/indicator or coder in the house that can do this?

What is your GARCH Oscillator supposed to display? The two coefficients or some forecast? And what the NN will try to find?

I suggest it should be similar to a RSI or stochastic oscillator. So one can backtest and/or estimate volatility (or standard deviations) and develop a forecast strategy wrt particular pairs.

NN (as has been the case) gives reasonably accurate forecasts in Trends. it just fails mostly in ranging periods

So in summary, one can make decisions from NN, having first ratified that price is not ranging (with Garch).

Hello! After a few months of work did you come up with something interresting? I'd like to give another try to the AR(F)(I)MA-(*)GARCH models but I'm always blocked on the huge number of samples required (like 10000 for ARFIMA). I'm interested in mesuring not forecasting. My wavelet decomposition gives me bad results when the volatily suddenly varies; then it takes some time (always too long!) to converge again. I'd like to see if I can combine the two methods (don't know how yet).

Hi to all,

I would like to share with you what I am doing using ARMA model and the residuals analysis of the model.

I am working on EURUSD, and I am trying to find a way to place order based on ARMA modelling.

Collecting Data, Data transformation, and Model fitting:

I am collecting each EOD Close of EURUSD instrument, and I calculate the log differencing to transform this time series in a stationary process. Then, using Box Jenkins, I fit the parameters of the ARMA Model.







Residuals Analysis:

After the model fitted, I analyze the residuals of the model. The process of the model residuals is a stationary process and follows a normal distribution.

Attached Image (click to enlarge)

Attached Image

Attached Image (click to enlarge)

Trading Strategy:

Like the model residuals is a normal distribution, I calculate the cumulative probability using the following formula:

Attached Image

Then I am able to display on the chart with a minor timeframe H4 or H1 what will be the tomorrow position with their respecting probability based on volatility max of 2 standards deviations:

Attached Image (click to enlarge)

Is that a good way to work with the ARMA model? What do you think about this strategy? Can we improve it together?

Thanks

David