Before defining what spatio-temporal forecasting is, it is best to define what spatio-temporal means and what forecasting is.
 
<h4>Spatio-temporal</h4>
 
Basically, spatial refers to space and temporal refers to time. In data science, spatial temporal, or spatio-temporal term is used when the data has both space-related properties (the spatial dimension) and time-related properties (the time dimension). As an example, energy consumption of a city can be classified as spatio-temportal. Such data is expected to have time dependent behavior such as monthly/daily/hourly at different districts of the city. Hence, each row of this data will correspond to a certain time-frame (the time dimension) and to a specific region of the city (the space dimension), yielding a spatio-temporal structure. For example, one row of this data may indicate the amount of energy consumed in the Midtown area of Manhattan, NYC on the 15th day of December 2021. Granularity for space and time may be higher or lower depending on the data set.
 
<h4>Forecasting</h4>
 
In dictionary, forecasting is defined as a statement about what someone thinks will happen in the future. In data science, forecasting is defined as estimating the future events based on the past and present data, most commonly, by analyzing the trends and fitting a formula that gets close to the data as much as possible.
 
Combining the two definitions, spatio-temporal forecasting is about estimating the type of event that will occur at a specific time and location in the future. For the energy consumption example, estimating how much energy will be consumed tomorrow in the Midtown area of Manhattan, NYC is considered a spatio-temporal forecasting.
 
For producing spatio-temporal forecasts using data science approaches, past data need to be analyzed and appropriate mathematical models should to be generated, which can define past events as a function of spatial and temporal properties of the data. For example, if we try to predict the energy consumption at a specific district for the next weekend, then we need to find a mathematical model that outputs the amount of energy consumed as a function of time-of-day, day-of-month, month-of-year, temperature, demographic information, economic indicators for that district etc.
 
Finding the mathematical model that will produce the most accurate results for a spatio-temporal forecasting problem can be a simple task in certain cases, but in general this is a very challenging goal to achieve. Especially for cases where expert knowledge exists and can be used to produce reasonable forecasts (for example the expert at an energy provider company can predict with a reasonable deviation how much energy will be consumed tomorrow without using any mathematical formula), custom deep networks are required to find the best mathematical model that will surpass the forecasts of the expert. Based on all our expertise in this field, we believe that <a href="/blog/why-do-we-think-data-science-is-an-art">this is art</a> and requires artistic mastery besides high quality engineering.

In dictionary terms, prediction is a statement about what someone thinks will happen in the future. In data science terms, prediction is estimating future events based on past and present data most commonly by analyzing the trends and fitting a formula that gets as close to the data space as possible.

What is spatio-temporal forecasting?

Predy is a spatio-temporal forecasting system that is built on top of a very powerful framework. This framework has been developed by the Predy team to serve as the underlying platform on which the end users forecasting system could be built on. Our spatio-temporal forecasting framework has unique and very powerful features distinguishing it from its peers.
 
The Predy Framework can operate on various data forms in terms of dimension and representation:
 
<ul>
<li>Dimension: The Predy Framework can readily process temporal, spatial, spatio-temporal and graph-structured data</li>
<li>Representation: The Predy Framework can process discrete (grids with boundaries) and continuous (points with timestamps) data</li>
</ul>
 
The Predy Framework can combine in highly novel ways different forecasting models both in time and space to optimize the final prediction outcome including:
 
<ul>
<li>Statistical models: conventional methods such as ARIMAX, ETS, and similar.</li>
<li>Classical machine learning models such as decision tree based models (XGBoost, LightGBM, random forests), MLP, SVR, GMR.</li>
<li>Deep models such as RNN (LSTM and GRU), TCN, CNN.</li>
<li>Point process models such as spatio-temporal Hawkes process.</li>
<li>And variants and hybrids of these models such as improved versions of the methods employed in different forecasting competitions, M3, M4, M5, ASHRAE.</li>
</ul>
 
The Predy Framework can utilize exogenous variables (side information) in different contexts such as:
 
<ul>
<li>Static side information: geographic data</li>
<li>Periodic side information: day/month/year data, special dates</li>
<li>Predictable side information: weather data, demographic data</li>
<li>Unpredictable side information: news, social media, other related problem-specific time series data</li>
<li>Sequence dependent features: mean and variance over the last week, month, year</li>
</ul>
 
Predy Framework supports online learning, which enables the system to utilize newly arriving data to constantly follow the trends, changing statistics, fine tune and improve itself continuously.
 
Predy Framework can process huge amounts of stored/streaming data without significant memory constraints, and it internally versions preprocessed data to decrease overhead if desired.
 
Predy Framework utilizes a sophisticated ensembling mechanism designed by our expert data scientists to mix different expert algorithms to optimize the final forecasts. Ensembling is performed in four layers:
 
<ul>
<li>Hierarchical level: combines the forecasts produced in different spatio-temporal resolutions</li>
<li>Regional level: combines the forecasts produced for different spatial regions / graph nodes</li>
<li>Dataset level: combines the forecasts produced by different expert algorithms trained on different subsets of the data</li>
<li>Combination level: produces the final forecast by ensembling the forecasts of different meta-experts based on all the related side-information</li>
</ul>
 
Predy Framework can make short/long-term forecasts ranging from microseconds to decades ahead depending on your needs. It can concurrently produce predictions in different time resolutions.
 
Predy Framework can manually or automatically cluster multiple time-series (multi-variate or spatio- temporal time series) into subsets based on numerous clustering methods to perform regional level ensembling.
 
Predy Framework provides a feature extraction module, which can extract a wide range of generic and custom features from your data to be used by our or your models.
 
Predy Framework contains customized utility modules to ease the development and deployment stages such as custom logging module, configuration generation module, evaluation module, internal data processing modules such as custom data/forecast normalization/scaling modules
 
Predy Framework uses MLFlow to handle experiment versioning and monitoring.
 
Predy Framework employs distributed training.
 
Predy Framework provides interactive UI tools for data analysis, data visualization, config generation, experiment and result monitoring.
 
All functionalities in Predy Framework are highly modular and scalable. It is possible to use it as a very basic time series forecasting model or to produce predictions for all the goods in a store without any overhead.
 
Thanks to all these capabilities of the Predy Framework, we are able to try many alternatives to find the best forecasting model for your specific needs, and furthermore, we can do it in a very short time.

Predy is a spatio-temporal prediction system which is built on top of a very powerful framework.

What is the "Predy Spatio-Temporal Forecasting Framework" about?

What we are trying to achieve at Predy is building custom and highly accurate prediction models for our customers so that they know in advance what is coming next. When our customers have this information early enough, they have the chance to be prepared and maximize their profits. Having highly accurate prediction models for your business not only provides you the opportunity to be prepared for the most obvious and probable cases, but also gives you the power to consider all probable cases. This is what we call what-if scenarios.
 
What-if scenarios refer to the case where one considers all contingencies and try to find the optimum action plan to cover them all. Of course, this is not an easy task especially if you do not have the right tools. The prediction models that we produce for our customers are perfect for this purpose and give our customers the power they need to be prepared even for unlikely events.
 
When we build a prediction model for our customer, what we do in basic terms is to find the optimum mathematical formula that models the target output in terms of all the related direct and indirect inputs. Take the prediction model of the sales figures of a retail market as an example. Such a model will define the amount of sales in terms of all the related input data such as past sales figures, date and time, GDP, location of the branch, weather conditions, competitor prices, trends, ad spend, seasonality etc. Once the prediction model is built, it is possible to predict what the sales figures will be for the coming days in terms of the values of all these parameters.
 
While some of these parameters are known when the prediction is made, some of them may not be known precisely. For example, when trying to predict the sales figures of the upcoming week, you need to input the weather forecasts of the upcoming days into the sales prediction model. Sales figures may be radically affected for certain items depending on the current weather conditions or campaigns that you are planning. However, once you have a trustable prediction model, you can set all the input parameters as you like, and obtain corresponding predictions for all possible (what-if) cases. If your prediction model is accurate enough, as we provide, then you have a very important tool to simulate the future and plan for it. This will give you the power to be prepared for all scenarios, avoid surprises and maximize your profit.
 
What-if scenarios, in this regard, are important use-cases for Predy and our customers benefit a lot from it.

What-if scenarios refer to the case where one considers all contingencies and try to find the optimum action plan to cover them all.

A powerful use-case for spatio-temporal prediction: What-if scenarios.

Dr. Jabe Wilson says, “Data scientist is like an artist working on a canvas…” and We strongly agree with him!
 
The data scientist works continuously on a problem and progressively develops alternative solutions, leading to a greater understanding of the problem. Likely, the process of an artist working on a painting over several weeks and seeing the canvas changes towards his/her vision is similar. In this blog, I will reveal the process behind this art and how data scientists solve problems and take steps of creativity.
 
We can summarize the steps taken by the data scientists for solving a problem as:
 
<ul>
<li>Problem definition</li>
<li>Data creation and feature engineering</li>
<li>Algorithmic design</li>
<li>Adjusting the whole system</li>
</ul>
 
The problem definition is essential, and it is the part where the math is highly involved. In general, you would build your hypothesis here, and then mathematically formulate it. This part shapes your algorithm choice and all the design afterwards. For this, you should find answers to questions such as “how can I formulate the business goals as measurable and optimizable performance metrics?”, “what is my possible inputs and desired outputs?", "is the problem supervised or unsupervised?", "is my problem classification or regression?", "is there a spatial, temporal, categorical relation in the data?" Your answers to these questions define your problem and shape your overall design. Sadly, answering all these questions in a concrete way is not always easy, however, this is the most fun part of it.
 
Data creation and feature engineering are where you decide your features and explore your data. Although the promise of deep learning is to achieve this in an end to end manner, it is not always easy with limited data. This process includes many preprocessing steps: removing noise and outliers, reducing dimensions, performing scaling and similar. The observations you had in these parts will also shape the performance and design of your algorithm. For example, you could explore that your target values are significantly correlated with one of the features. Still, this correlation could mean nothing because “correlation does not mean causation”. If, in such a case, your model wrongly interprets this relation, then it may perform poorly.
 
The algorithmic choice is greatly dependent on the previous steps and it is where the ingenuity happens. Considering the algorithm as a black box that tells the truth is mostly a misconception. Methods telling you up to a certain level what is happening inside this black box, such as in deep learning models, are in early stages of research. Thus, “explainability” is required. Furthermore, most complex algorithms usually fail in real-life applications due overfitting since most of the time you won’t have enough data for modelling or the model changes constantly. It is possible to obtain high scores by selecting algorithms with trial and error. However, this process is costly or even impossible, and you could waste valuable time. In every piece of the design choice, you must ask "why" to create awareness of the process and algorithms.
 
Also, you need to have extensive knowledge of your algorithms including all their parameters, weaknesses, advantages since you need to tune them to avoid overfitting or underfitting. Monitoring your algorithm's learning process and performance with the proper metrics allows complete control of the system. Thus, a data scientist should also have a great command on monitoring tools and metrics.
 
Lastly, following the latest development in the literature is a must. Data science is a vast and dynamic area. Nearly every day, new algorithms, methods, tools and results are introduced in different domains, which may greatly change your understanding, approach, solution and even the problem you are working on.
 
Summing up all together, data science is a field that requires expertise, sensitivity and artistic mastery at every stage. This is even more true for operationally used artificial intelligence systems. At Predy, we aim to produce predictions for your future needs, which is normally a task carried out by expert staff or conventional software systems. In order to meet and exceed your performance expectations, we have little or no room for imperfections at any stage. That is why, we believe that “data science is art” and we define our values as “Meticulousness”, “Patience” and “Creativity”, i.e., the skills that a top data science artist should have.
 
We are meticulous; we love data and we study every detail of it meticulously.
 
We have the patience; we aim for the best and strive for it to the limits.
 
Our team is a collection of creative people; we think out of the box and consider every possible option.

Data science is a field that requires expertise, sensitivity and artistic mastery at every stage.

Learn more about the techonology behind Predy

Ready to get started?

Get Started Now

Let us keep you updated on Predy!

Company

Links