0
$\begingroup$

Currently, I found the right recipe for a time series regression problem to finally get acceptable to good results.

Here is the config file

{
    "data": {
        "sequence_length":45,
        "train_test_split": 0.85,
        "normalise": false,
        "num_steps": 10
    },
    "training": {
        "epochs":30,
        "batch_size": 32
    },
    "model": {
        "loss": "mse",
        "optimizer": "adam",
        "layers": [
            {
                "type": "lstm",
                "neurons": 161,
                "input_timesteps": 45,
                "input_dim": 161,
                "return_seq": true,
                "activation": "relu"
            },
            {
                "type": "dropout",
                "rate": 0.1
            },
            {
                "type": "lstm",
                "neurons": 161,
                "activation": "relu",
                "return_seq": false
            },
            {
                "type": "dense",
                "neurons": 128,
                "activation": "relu"
            },
            {
                "type": "dense",
                "neurons": 1,
                "activation": "linear"
            }
        ]
    }
}

Here is the results I got

enter image description here

What can be good improvements I can bring to my model so that I can get better results? There are a lot of small spikes and other places where the curve is rather constant that I should get a rise or fall of the curve.

$\endgroup$
0
1
$\begingroup$

Try using an RMSProp optimizer instead of Adam optimizer. Also try decreasing the batch size and keep a small learning rate like 0.001.

$\endgroup$
1
  • 1
    $\begingroup$ This is for a time-series forecasting problem, so yes, LSTM is useful here. $\endgroup$ Nov 7 '18 at 1:58

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.