Validation is what you use to answer your question. See the attached for references. Best wishes David Booth

This is very ambitious. It is very dependent on your dataset, problem formulation. If you want to know the basic concept, I recommend the book Forecasting: Principles and Practice (https://otexts.com/fpp2/). There is a chapter to introduce some fundamental ideas of machine learning and the simple NN method. By the way, you also check some updated published papers in the past 2 years in top conferences or journals in a certain area.

I believe it is critical to narrow the scope. After all, there is no 'Best Algorithm ' in the general areas. For example, If you are interested in electricity price forecasting, Applied Energy, Energy, or IEEE Trans could be good options to take a look at first.

Hope this helps.

Deep Learning and Machine learning have several techniques for time series forecasting. We cannot come to a conclusion that one particular technique is the best technique. It depends on several factors such as the type of distribution of the dataset(whether it is normal or non-normal), presence of outliers, trend, cyclic, seasonal patterns in the data. Kindly refer to the book "Deep Learning for Time Series Forecasting" by Jason Brownlee. It's an excellent book where they mentioned step by step process in choosing a suitable technique for a particular data using Python

@David Eugene Booth thank you to your kind information.

Dear Ashifur Rahman

Apart from your data and its characteristics,

You also need to finalize the tool/software you are compatible with

Python is open source and MATLAB is paid. (One month student free usage is allowed; Check the website for more details)

As suggested by previous answers, you may then explore various algorithms.

Regards,

Dear Dr Ashifur Rahman . Autoregressive Integrated Moving Average (ARIMA): Auto Regressive Integrated Moving Average, ARIMA, models are among the most widely used approaches for time series forecasting. See the link: https://www.advancinganalytics.co.uk/blog/2021/06/22/10-incredibly-useful-time-series-forecasting-algorithms

However, Recurrent Neural Networks are the most popular Deep Learning technique for Time Series Forecasting since they allow to make reliable predictions on time series in many different problems. The main problem with RNNs is that they suffer from the vanishing gradient problem when applied to long sequences. See the link: https://towardsdatascience.com/time-series-forecasting-with-deep-learning-and-attention-mechanism-2d001fc871fc

To compare see the following useful link: https://www.mygreatlearning.com/blog/deep-learning-vs-machine-learning/

There is no "best tool" in analysis of this kind.

You have to check which is the most appropriate means, based on the data you have and the objectives you set yourself.

I have read that neural networks tend to perform well when using time series data

An Experimental Review on Deep Learning Architectures for Time Series Forecasting:

https://arxiv.org/abs/2103.12057

Hi,Ashifur Rahman

Please see these references

Deep Learning-Based Time Series Forecasting, DOI:10.1007/978-981-16-3357-7_6, In book: Deep Learning Applications, Volume 3.

Deep Learning Approaches to Time Series Forecasting, DOI:10.1201/9781003102281-6, In book: Recent Advances in Time Series Forecasting.

Time-series forecasting with deep learning: a survey, Philosophical Transactions of The Royal Society A Mathematical Physical and Engineering Sciences 379(2194):20200209, DOI:10.1098/rsta.2020.0209.

GOOD LUCK

Ashifur Rahman

For time series forecasting you can use neural network. RNNs are deep learning model used for solving problems based on sequential input data such as time series.

ETS (Exponential Smoothing)

Exponential Smoothing or ETS algorithm is one of the easiest and fastest algorithm to forecast quite accurately time series. It’s capable to handle trends and seasonality and is easy to interpret.

It can be implemented with one single line of code in R or Python and also comes embedded in tools such as Alteryx. The AAA version of it is even embedded in Excel as the FORECAST.ETS function and in PowerBI to forecast Line charts.

Whatever your time serie problem is, it’s always good baseline to start with.

ARIMA / SARIMAX

Just like ETS, ARIMA / SARIMAX are part of the old yet very good Forecasting Methods for Time Series.

It also provides a very good baseline and is easy to implement using a single line in R or Python. It’s also embedded in Alteryx’s Desktop.

For Python implementation of ETS and ARIMA models, you can use the statsmodel package.

Prophet

Facebook developed its own Time Serie algorithm in 2017: Prophet.

Very easy to implement with a few lines of Python or R, it provides a forecast which is easy to interpret, the algorithm not being overly complicated. Compared with ETS it is capable to deal with changing trend patterns and is better tailored for high frequency (daily or more) data points. Users can also incorporate some business intelligence by setting floors and ceilings to the forecast, and Prophet uses this information to dampen the trend accordingly if need be.

Random Forest and Gradient Boosting (Machine Learning)

Machine Learning has taken off a lot in the past years and provide great possibilities for forecasting time series at scale. Such algorithms can process both historical time series inputs but also external relevant features which can increase accuracy. Among them Random Forest and Gradient Boosting are some of the most performing, all running exteremly fast and providing excellent performance

RNN and LSTM (Deep Learning)

Deep Learning also provides interesting methods to forecast Time Series. Among them Recurrent Neural Networks (RNN) and LSTM cells (Long Short-Term Memory) are popular and can also be implemented with a few lines of code using Keras for example.

N-BEATS

N-BEATS is a custom Deep Learning algorithm which is based on backward and foward residual links. It is popular among forecasting competitions, outperforming past winners of M3 and M4 competitions. It was also widely used among the best performers of the M5 competition in summer 2020. It presents many benefits, especially being interpretable and quick to train

It is available in the PyTorch-Forecasting package.

DeepAR and Temporal Fusion Transformer (TFT)

To further increase Forecast Accuracy it’s now common to combine historical data points with external information such as (price, discount, temperature, population…) which can help the algorithm better understand the context that led to a particular time series data point.

The Machine Learning (ML) and Deep Learning Algorithms mentionned above (Linear Regression, Random Forest, Gradient Boosting, Deep Neural Networks…) are all capable of dealing with these extra features, however Amazon and Google have both developed their own dedicated custom algorithms when it comes to time series: DeepAR (Amazon), available on Amazon Forecast (AWS) and Temporal Fusion Transformer (Google), which is available in the PyTorch-Forecasting package.

I think this article can add a great idea for algorithms for TS forecasting but in practice the econometric model is better than others. Using arima sarima arch garch than a stochastic who overestimate or underestimate values.

Thanks everyone.

Ashifur Rahman

Genetic Programming (GP) can also do that. Basically, with GP, you can discover a mathematical function (that contains +, -, *, sin, log, constants, etc) that has window_size arity (that parameter is set by you) and whose output is the next value to predict.

Attached is the result of a search through open-access books using "Time Series" search in the title.

Variants of Transformers

Recurrent Neural Networks for better suitable for your problem amongst deep learning algorithms.