The Use of Artificial Intelligence Models for the Prediction of Temporal Data
For several years, there have been numerous techniques for forecasting temporal data, some of which are very simple. Others, such as neural networks, are much more complex. To apply these artificial intelligence models effectively, the data sets used must be well-cleaned and prepared beforehand in order to provide the models with the most qualitative data possible.
Indeed, most AI models are not designed to handle missing or poorly structured values. In the first part of this article, you learn about the different data preparation methods that optimize performance of future models.
When it comes to modeling, we can group the forecasting methods into four main categories. In each of them, there are many variants:
- Simple Moving Average (SMA)
- Simple Exponential Smoothing (SES)
- Autoregressive Integrated Moving Average (ARIMA)
- Neural Network (NN) based methods such as LSTM
Simple Moving Average (SMA)
The simple moving average is the most basic method of forecasting. It is an average of a subset of periods in a time series. Moving averages are usually plotted as a line graph to give an idea of the overall trend of the series. They can be useful for confirming the direction of a trend, or getting a visual of its magnitude.
SMA is easy to understand, and often preferred to more rigorous statistical methods. It gives a good view of the trend, and smooths out short-term fluctuations. It also reduces the effects of extreme values. However, SMA does not have a statistical methodology for determining the forecast period.
Simple Exponential Smoothing (SES)
This is another popular technique for smoothing time series data. The moving average is a simple average where all observations are applied with equal weight. The exponential average assigns decreasing weights over time. For example, the most recent observation should have slightly more weight than the second most recent and so on.
SES is generally used to make short-term forecasts. It is more efficient than SMA, because it gives a higher weight to more recent data points than the equal weight given by SMA.
However, SES cannot reliably make longer-term forecasts, mainly because this method does not take into account any trends in the data. Therefore, extensions of the SES model such as Holt’s double smoothing and Holt and Winter’s triple smoothing will need to be examined.
Autoregressive Integrated Moving Average (ARIMA)
The previous models lead us to new forecasting techniques using neural networks to overcome some of the problems associated with more traditional methods.
ARIMA has long been a standard method for forecasting time series data. Although ARIMA models are widely used in modeling economic and financial time series data, they have major limitations. For example, in a simple ARIMA model, it is difficult to model non-linear relationships between variables.
Recurrent Neural Networks – Long-Short Term Memories (LSTM)
Les modèles précédents nous conduisent à de nouvelles techniques de prévision utilisant des réseaux de neurones afin de surmonter certains des problèmes liés aux méthodes plus classiques.
LSTM models are based on recurrent neural networks, a sub-family of artificial intelligence called Deep Learning. They can be used to predict temporal data (as well as other recurrent neural networks). The state of an LSTM network is represented by a state space vector. This technique makes it possible to keep track of the dependencies of new observations with previous ones (even very distant ones).
In general, LSTM models are complex and are rarely used to predict a single time series, as they require a large amount of data to be estimated. However, they are commonly used when predictions are needed for a large amount of time data, and offer remarkable performance.
Challenges in Time Series Forecasting
We would like to share our experience in conducting time series forecasting projects, and identify challenges that the development team may face.
Lack of Data
The larger the datasets, the more training data the system can access, which leads to greater accuracy in predictions. However, there are some limitations to consider related to a lack of historical or seasonal data for a target variable. Therefore, depending on the model used, a lack of data could lead to an overall decrease in prediction accuracy. This is referred to as under-fitting.
Lack of Knowledge in the Target Field
Knowledge about the field can help improve the quality of the models in any project. To avoid a knowledge deficit in the domain, the involvement and expertise of business niche specialists is required.
Using the Wrong Model
It is essential to use a model that is adapted to the data, to the volume and to the type and duration of the forecast, whether short, medium or long term. Indeed, the use of a model that is too complex could lead to an under-adjustment of the data. On the other hand, a simpler model such as SMA or SES will probably not give any results on data with seasonal or cyclical variations.
In Summary
In these two articles, we have explored time series forecasting and its most important elements, i.e. the constituent components into which a time series can be broken down when performing an analysis (trend, seasonality etc.).
We have also examined different types of forecasts and looked at moving averages, ARIMA and neural networks.
The complexity of implementing a time series forecasting project requires the highest quality of development, which our experts can provide.
With the knowledge that Uzinakod has accumulated in similar projects, we will certainly meet the requirements of your project in terms of thorough consideration of the specifics of the domain, from business objectives to the design of the various forecasting models. Do not hesitate to contact us to talk about your project!
