Help with ANFIS time series prediction

Hello,

I am trying to use ANFIS to perform time predictions on some data. More precisely, I want to make prediction on a set of data based on the past values of the same data. I first generate my time series based on the original data and the delays I want to use:

load('TrainingData.mat'); % Vector 'x' is loaded with original data

Delays = [1 2 3 4 5 6 7 8 9];
DelayNum = length(Delays);

[Inputs, Targets] = CreateTimeSeriesData(x, Delays);
Inputs = Inputs';
Targets = Targets';

The "Inputs" vector contains [x(t-1), x(t-2), x(t-3) ... x(t-9)] on each column, according to the delays vector I chose, for t from 0 to N, where N is the size of x. "Targets" is a vector composed by x(t) with t = 0 to N again. I then divide the time series into two: one for training (TrainInputs, TrainTargets) and another one for validation (ValInputs, ValTargets). Next step is generating the FIS using genfis3 and train the system with ANFIS passing some additional options:

fis = genfis3(TrainInputs, TrainTargets);

opt = anfisOptions;
opt.ValidationData = [ValInputs ValTargets];
opt.InitialFIS = fis;          

fis = anfis([TrainInputs TrainTargets],opt);

The reason I am using genfis3 is that every time I try and use genfis, MATLAB gives me the ever feared "Loop is too big" error and proceeds to eat up all of my RAM to the point where I have to force-reset the computer.

Now, prediction is the last and most tricky step. Its the one I am in doubt about. After doing error calculations and plots to check if FIS training performance is good, which it seems to achieve almost everytime, I generate a new time series based on another set of data, PredictionData:

load('PredictionData.mat');
Limit = 150; %This is the number of points where I assume data generation for the "PredictionData" set has stopped. I have to predict the rest of the data (which is until data point number 315).

xPred = x(1:Limit);
xPredTotal = x;

[PredInputs, PredTargets] = CreateTimeSeriesData(xPred, Delays);
PredInputs = PredInputs';
PredTargets = PredTargets';

Lastly, I use an evalfis loop to predict a future value for the inputs, and feedback the predicted value into evalfis again, and so on:

PredSteps = 315 - Limit;
PredInput = PredInputs(length(PredInputs),:);
PredOutputs = zeros(1, PredSteps);
for i = 1:PredSteps
    PredOutputs(i) = evalfis(PredInput,fis);
    PredInput = ProcessInput(PredInput, PredOutputs(i));
end

The results unfortunately don't look very good:

By looking at the final results, what I imagine is that the FIS prediction (in red) is just following the training data pattern (in cyan) with a time shift of ~80 points late. My first guess, of course, is that overfit is happening. That is why I added the whole validation data thing, but that almost didn't make any difference. In the help and doc sections of anfis, evalfis and anfisOptions, that is the only hint I found about reducing the chance of overfitting in ANFIS, so it seems I ran out of options.

So my questions are: is the prediction process correct? In case it is and the results I am seeing are the ones the FIS is supposed to produce, what can I do to make it a little more "reactive" and more responsive to prediction data?

Thanks,

Jonas