Unable to predict data well enough

Hi,

I have a dataset of 19 non linear regression data points. I am training the first 10 and trying to predict the next 9 values using different Neural Network. However, after trying out mulitple Neural Networks such as Radial Basis Function, Baysian Reguralization BackPropogation, Function Fitting Neural Network and LSTM, i am still not getting good prediction results. I have attached the data where X is the input and Y is the output. First 10 data points are being used for training and next 9 points are used for testing. I have also attached by code which uses Baysian Regularization backpropogation.

Attached is the dataset link :

Data File Link

The results obtained using Baysian Regularization backpropogation are shown in graph attached below :

The code used is as follows:

close all;

clear all;

rng('default');

out_col = 2;

inp_col = 1;

data= xlsread('Neww_BRP.xlsx');

n = 7;

Neurons = 5;

X_Train=data(i:n,inp_col);

Y_Train=data(i:n,out_col);

XValidation = data(n+1,inp_col);

net = feedforwardnet(Neurons,'trainbr');

[net,tr] = train(net,X_Train',Y_Train');

y = net(XValidation')';

Kindly let me know how can i improve the prediction results?

Thank you.

Answer 1

[ME]

I am far from an expert in this area but I'd guess that you are struggling to get a good preciction because you don't have enough data points to train your network properly.

[Gulrukh Turabee]

Hi,

thank you for your answer.I have seen some research papers, where only 6 to 7 data points are being taken for training, still they are getting good prediction results. I doubt this is the issue.

[ME]

Have you tried using a random sample of you data points to train the network rather than just using the first ten? That way your training data will likely cover a greater range of the behaviour seen in your data set and perhaps improve the predictive ability of your trained networks.

Answer 2

[John D'Errico]

I would point out that your code shows you using the first SEVEN data points, not the first 10, despite your claim otherwise.

n = 7; 

As well, since those first two points are completely inconsistent with the rest of your data, I'd expect to see a serious problem in any intelligent long term extrapolation. You gave it 7 data points, and 28% of your data was completely useless crapola. Just because someone else had success does not mean that your data is as good as theirs.

I would instead, suggest that you really try using the first 10 data points. Better yet, try using points 3:10 to train the net. Then see how well prediction actually proceeds for points 11-19.

Remember that extrapolation is always a risky business, prone to failure. If it was always so easy to do, then we would always have perfectly accurate weather forecasts.