# Very simple real-valued time-series dataset for RNN prototyping

Is there a simple real-valued time-series dataset on which a vanilla RNN model can be trained. With "very simple" I mean only two to four real-valued inputs per time step and a single real-valued output per time-step.

Background: I am doing research in the intersection of machine learning and formal methods. To test a new technique for formally verifying RNNs, we need to start with a quite simple setup.

Thanks!

Another way would be simulate a time series, with any characteristic that you want. A very simple example in python:

import numpy as np

X = np.array([np.arange(0, 100)])
y = np.array([np.sin(np.arange(0, 100) * np.pi/4) + np.random.uniform(0, 1e-2, 100) + np.arange(0, 100)])

Simulating you're able to adjust tendency, seasonality, covariates and another feature of interest. And maybe this would be valuable in you research analyzing how time series features impact the performance both in RNN and formal methods.

You can take a look at some of these datasets in UCI:

These are quite straight forward, although they might contain more columns that 4 or 5. But they are quite simple.

Or you can take a look at this:

https://www.kaggle.com/uciml/electric-power-consumption-data-set

This is also based on UCI data.

Overall you can take a look at datasets available on Kaggle, you can probably find something there.

https://www.kaggle.com/datasets

Hope this helps.

These are my favourite time series datasets.

These two are quoted a lot in the literature on the imputation of missing data in time series:

If you are not interested in fighting with missing values (yet), I suggets you to take a look at the first two. There's plenty of work to do on all though, either univariate or multivariate analyses. You can also use Kaggle kernels and the literature as a benchmark to evaluate your own results. Good work, have fun.