I am looking for a Machine Learning tool for Time series analysis.
I have tried MATLAB (ANFIS tool) and ANN in R, but the error is more than my expectation in testing step. Furthermore, MATLAB process time is strongly dependent on the input layers and that doesn't allow me to use all of my simulated data as input.
Dear Javad,
The best machine learning tool is case dependent. It depends on your dataset. You cant choose one unless you try plenty of them. Generally preproccesing plays an important role in such tasks. The following link may help.
https://aaai.org/ocs/index.php/FLAIRS/FLAIRS17/paper/download/15475/14995
Hi Javad,
Actually, when it comes to studying/predicting the trends in Time Series data, regression based machine learning approaches are mostly applied. However, deciding the optimal ML algorithm is dependent a multiple factors which may include the structure of the Time-Series dataset you are considering in your research, the data cleaning/preprocessing technique applied, and the robustness of the ML algorithm. In principle, SVM Regression models have been shown to be robust and at the same time provide good estimation results in comparison to other ML algorithms. Since you have tried ANFIS toolbox in MATLAB and NN in R with unsatisfactory results, I would suggest you possibly try SVM Regression.
You may also check out the following URL for more information.
https://www.datascience.com/blog/time-series-forecasting-machine-learning-differences
Hope you would find this tips helpful in your research.
Wish you good luck.
Machine learning is a fantastic way to evaluate time series.
There are a number of approaches that can be implemented. Some are better than others depending on the characteristics of the time series. Methods include:
- Regression - Using time-based features such as week, month, day, day of week, etc as predictors. You can also add in external predictors that may influence the target (e.g. weather and temperature may affect sales of umbrellas).
- ARIMA - Autoregressive Integrated Moving Average - Using autocorrelation (lags) as predictors
- GARCH - Models changing variance
- Time series decomposition - Splitting out trend, seasonality, etc
- Others - Deep Learning, GAM-based models (prophet) - Can also be useful
Here are a few articles that can help, which show different machine learning regression techniques on a data set with high presence of monthly seasonality and trend:
- H2O for time series - Uses H2O automatic machine learning algorithm in R
- Deep Learning with Keras for time series - Uses the Keras library for forecasting the Sunspots data set in R.
- Multiple regression for time series - Uses the lm() function in R
- ARIMA for time series - Uses the forecast package, auto.arima()
Machine learning (using decision trees, linear regression, etc) may not be well-suited if patterns are non-existent in the data. In some situations other time series-specific methods may be better:
- ARIMA is great when autocorrelation is present, but possibly less pattern such that Machine Learning might have an issue picking up a relationship. Great for random walk-style series.
- GARCH Like ARIMA but is better suited where variance changes (e.g. shock like financial crisis). Great for random walk-style series.
Last, for those that are new to time series, here are a collection of posts that help define rolling functions and autocorrelation:
- Lags and Autocorrelation - Detecting relationships between the time series and previous versions of itself
- Rolling Correlation - Useful in detecting relationships between two time series
- Rolling Functions - Rolling mean, median, quantiles, etc
- Period Apply - Applying functions by time period typically an aggregating function (mean, median, sum, etc)
Greetings and Regards
I think MATLAB software is a good option.
- Badges
- Science topic