As illustrated across our website, time series are everywhere. Each industry vertical, each domain, each company comes into contact with time series in one way or another. In the use case library below, you can explore the extensive applications of time series, or look for applications in your specific field of interest.
Michal Bezak - Tangent Works
Contact centers typically operate with a pool of resources. For contact centers, predicting volume of incoming requests at specific times is critical to make proper resource scheduling. In such case, forecasts are expected for very short, and short-term, horizon (e.g. a week ahead).
High-quality short-term forecast brings confidence that FTEs (full time equivalent) planned for the next week are just right for delivering on SLAs, not to mention other benefits, such as higher confidence when planning absence, or improving morale of employees who would not face overload from “sudden” volume peaks.
Predicting volume of requests for mid-term horizons, e.g. 3-months ahead for weekly data, is important input to resource management. It takes time (weeks if not longer) moving people around, hiring, upskilling, or down-sizing pool of resources. Because of this, forecasts for longer horizon are needed, starting from one to more months.
The picture below depicts how contact centers are typically linked to WFM (work force management), internal departments and other factors. This provides intuition which factors should be included in data used for building models and forecasting.
Big contact centers would support not one, but multiple regions, cultures, and languages. Very likely forecast by language, or country would be beneficial.
Having forecast for just one perspective, one or two prediction horizons is not sufficient, also dynamics of ever-changing business means that using model built one month ago is suboptimal. This means that having capability to build models and make new predictions instantly is necessary for successful management of resources.
TIM can forecast for any prediction horizon, from intra-day to days or weeks ahead, and thus can be used for short term, midterm as well as long term forecasting. It can build new model in faction of time, building models from the latest data which helps to achieve better accuracy.
It’s automated, thus your analysts and data scientists have free capacity to focus on other agenda. You’d gain new capability, more frequent forecasting or forecasting per various perspectives with minimal additional effort is possible.
Effort and time required to set up forecasting is reduced to fraction of what would be typically required. TIM, by design, automates most of the steps required for set up and operations, and offers robust ML solution capable to quickly adapt to structural changes.
High-quality forecasts would deliver following benefits:
Explanatory variables should include historical actual volume values, meteo predictors, holiday information, marketing activity (campaign) information, factors describing customer base, planned outages, and/or other relevant data with as low latency as possible.
TIM’s output consists of forecasted volume of requests per each hour/day/week in selected prediction horizon.
For physical reasons, transport of electrical energy via electricity transmission grid leads to losses; thus, less power can be withdrawn than is fed into the grid. Losses are generated, for example, through resistance in lines or transformers, and are released as heat. Factors such as load, outside temperature and switching states in the grid have an impact on power loss. Transmission system operators (TSO) must compensate for losses, and thoroughly manage them because losses influence balance on the grid.
TIM proved to deliver highly accurate predictions for intra-day and day-ahead forecasts and is fully capable to forecast practically for any prediction horizons. Even more, effort and time required to set up such forecasting solution is reduced to fraction of what would be typically required. TIM, by design, automates most of the steps required for set up and operations, and offers robust ML solution capable to quickly adapt to structural changes.
Explanatory variables can include, besides historical actual values for losses, also technical information about relevant points on the power grid, load, and weather data.
TIM’s output consists of the forecasted active power loss in the same unit of measurement (typically kW or MW) and granularity as the input data, over the desired forecasting horizon.
Read the exciting success story for this Use Case.
How to prepare your retail and online production, inventory and distribution for changing COVID-19 measures?
External factors can have a huge impact on your demand forecast for specific products in specific locations and channels. The constantly changing government guidance during the current COVID pandemic can also cause huge swings in demand, especially when wide reaching regulation – such as decisions to close restaurants or limit movement – can come and go within hours for entire towns, cities or countries. To help businesses dynamically allocate their scarce resources of staff and inventory, the use of TIM and real-time instant ML can create forecasts that are as dynamic as the events that influence them. This can enable fast business decisions to take advantage of opportunities and limit the costs of reacting to current events.
Talk to our specialist about adaptive retail sales forecasting
The TIM Engine is perfectly suited for this application, because of the speed, resilience and ease of deployment. Other forecasting methods, such as statistical forecasting, are far too slow to react to the modern business climate. Single variate models will miss on the complex interaction between seasonal changes, externally driven changes to demand and externally driven changes to mobility. The speed of forecasting is also incredibly important in determining what factors are causing permanent changes to demand and customer behavior and which changes will be temporary.
Using the TIM Engine, an analyst can quickly iterate on models using dozens or even hundreds of features to get predicted impacts on demand in near real-time. This is a must-have tool for anyone in an organization who is responsible for planning of inventory and/or staffing levels.
As an example, using the TIM product we can immediately predict the impact a new restriction in a specific town would have on online ordering in specific postcodes. This can be used to ensure the correct allocation of capital equipment (trucks), staffing (drivers) and product (warehouses) even in advance of the restriction being implemented. With the TIM Engine, this forecast can be available in minutes to respond to significant events that may be happening within only a few days. We can see that a new announcement of a restriction to restaurants causes an immediate surge in online grocery demand that lasts for several days before subsiding to pre-restriction levels. This analysis can be extended to review the impact on specific products at a SKU level – including the ability to run independent demand forecasts for individual SKUs at individual stores in seconds.
Demand Data has real-time access to external data of significance, such as weather data (forecast and history) for any geographic point, as well as real-time COVID cases, mobility and restrictions for each postcode in most major countries. This data can instantly be prepared as inputs to be combined with sales data for specific products, channels or store locations. We have templates available which can plug into sales data at a store level and create the base models instantly (including COVID, weather and human mobility). From there, your analysts can iterate using other data or assumptions they have and get feedback in seconds on which assumptions or data are good predictors and which ones are not.
Talk to our specialist about adaptive retail sales forecasting
Watch a video taking you through this use case below:
Anyone who has ever sat in an S&OP meeting or in budget review will have heard this before: “We would have hit our target if it wasn’t for the (rain, wind, snow, sun, cold, hot…).” Very rarely are these comments backed up with real facts – but you will get the occasional nod around the room. Ah yes… I do recall it rained on the second Saturday in July – that must be why we missed our target by 12%. Or the line in a seasonal review meeting that says, “Our winter boot sales are up this year over last year because of the early winter storm this year and the late Indian summer in 2017, remember???”
The problem with these statements is that though they may be true, mostly they are easy excuses that cover true problems within the business. Maybe there are supply issues, or merchandising issues or a general decline or growth of our brand. Instead of comparing actuals to budget, what if a third line could be added called “Event adjusted budget”. This way, managers, planners and executives can understand the impact that events, such as weather, may have played on our performance, but most importantly, this will help us to understand and think about what else might be happening in our business. Sounds interesting?
The good news is that with advances in Machine Learning and tools like the TIM engine, this is entirely possible to do. By comparing actual sales history at different levels of your organization and bringing in weather conditions on the day as features, a company can start to understand the importance of these features on driving actual sales. Using this data, we can start to see what categories are heavily impacted (ice cream, umbrellas, or the heating bill) and what locations get impacted (outdoor shopping malls, online sales, restaurant delivery). As the TIM engine begins to understand what event variables are driving volatility, we can create historical forecasts based on simulated conditions. (This is what we would have predicted this year based on the same conditions as last year as an example.) This means that the next time your sales rep argues they would have made their target if it wasn’t for the cold front, your business can support this with real evidence. Not only that, but you can start to use live weather forecasts or long-term weather trends to drive predictive interventions – like adding an extra employee at your beach café this weekend.
With cloud storage being so cheap, detailed datasets containing years of history for hundreds of thousands of weather data points are readily available. Not to mention other event data like football games, traffic patterns, holidays and much more. This data is relatively easy to find and there are many services that make this data available historically and as forward-looking projections. With this data and your historical sales and historical budgets, the TIM engine can train models to produce forecasts based on simulated conditions (or after removing them entirely).
Watch a video taking you through this use case below:
Looking for a more in-depth view into this use case? Check out this video:
Launching a new product successfully can be a very stressful but extremely rewarding task if executed correctly. After weeks or months of planning, your new product or line extension is ready to go to market and in the initial few weeks or days of launch, speed and agility is critical. There are several complexities to consider with new product introductions (NPIs). First is that your initial forecast is a “cold start”, meaning that you must make many assumptions and usually have little data on how a product will perform. This makes it extremely difficult to anticipate how your product is going to be adopted in the market and how much inventory to buy or produce. Second is that usually, capacity is finite on an item that has not been made before and the inventory risk is very high – so rarely will your CFO want to take large inventory positions on an unproven item.
TIM’s InstantML can be an excellent solution to managing a new product forecast. Unlike other demand modeling tools, which require history to build models and then are inflexible to change models quickly, TIM can completely rebuild models as soon as the first set of sales data comes trickling in. With TIM, you can rebuild your models daily or even multiple times per day in the first several days of trading. Through storing the projections from each model, you can quickly evaluate how quickly the models converge or know immediately if your original projections are too high or too low. This enables the planning and launch team to react immediately and swiftly to rapidly changing conditions. It has been proven in the fashion industry that the market signal in the first few days of a new item trading is the single largest predictor of the overall long-term margin for that product. Isn’t this something you should be watching constantly and adapting to during those critical first few days?
A new product introduction will start with of course your initial forecast. If you have launched similar products in the past, you can use common features of the product to build an expected demand profile. In launching a new clothing item for example, you can use color, cut, fabric and price points as well as category to help build this. As we move into the launch phase, sales data is critical and getting as much detail as fast as possible is best. Sales by each transaction at each store or sales channel is ideal.
The output of TIM in this case is a timestamped forecast so you can see how the forecast is changing over the next several weeks hour by hour as new data comes in. Unique to the TIM product, is that every forecast can be built based on an entirely new model and each forecast can be compared to previous forecasts. When the forecasts converge and are consistent you can safely assume that demand has stabilized, however if they fluctuate or trend heavily up or down you can immediately adjust your production plan, distribution footprint or inventory policies.
The velocity of change in the mortgage industry is outpacing the abilities and reach of existing prepayment and profitability models, especially in the tail-risk coronavirus world. With over $16T total outstanding mortgages, the US mortgage market requires a scalable, accurate forecasting and modeling solution.
TIM enables users to iteratively forecast mortgage prepayment, delinquency, and default. This can help investors, GSE’s, servicers, lenders, and other stakeholders to evaluate and quantify the valuation and credit risk of their mortgage assets.
The US mortgage industry generates a vast quantity of data highly relevant to profitability and risk analysis. Some of this data is in users’ possession, some of it is publicly available, and some of it can be acquired. InstantML allows you to understand, differentiate, and quantify the relevance and impact of each data source to your forecasts.
Accurate forecasts of sales and demand in the retail industry can make a huge difference in allowing a company to adapt to changing times. Furthermore, sales planners frequently rely on forecast-driven tools to adjust levers such as product pricing and the timing of promotions. There is frequently also a need to apply these techniques dynamically across different levels of product hierarchy, geography, or other dimensions. However, many companies are still using relatively rudimentary forecasting techniques, which can affect the accuracy of these forecasts. Many enterprise-facing tools are also designed with inefficient workflows, reducing the ability to do effective analysis for end-users such as FP&A and sales planning teams.
Machine learning techniques provide the most up-to-date approach for accurate forecasting, but are time-consuming to implement from scratch. This is where TIM’s highly automated, lightning-fast, ML-driven capabilities can make the difference. Automation means that analysts can use the tool without needing experience in ML theory or in programming. Lightning-fast means that end-users will be easily able to apply forecasting to any data that requires it, and to quickly iterate on scenario planning. It also means that TIM’s forecasting abilities can be applied at scale, on any data that needs to be forecasted. It can also be used in interactive BI applications. For example, planning teams could use the same BI tools that they use every day, select a particular slice of data, adjust some inputs such as pricing and promotions, and use a TIM BI tool integration to get a forecast result – all without leaving the BI tool.
Typical inputs for a sales forecasting application might include historical sales (usually split across product hierarchy, geography etc.), regional store information, local demographics, level of competition, and indicators of consumer demand, industry performance, and economic performance. Sales planning applications might also include pricing and promotion start/end dates as adjustable inputs, allowing planners to see their effect on the sales forecast.
The recent evolution of Internet of Things (IoT) technologies has resulted in the deployment of massive numbers of sensors in various fields, including manufacturing, energy and utilities, and logistics. These sensors produce huge amounts of time series data, but understanding the data generated and finding meaningful patterns remains an obstacle to successful IoT implementations.
A common problem that can be solved with IoT data is anomaly detection, where temporal patterns in the data are used to identify mechanical or electronic anomalies in a piece of equipment, prior to the occurrence of a failure. This approach can help to minimize downtime for manufacturing pipelines or other IoT networks, thus preventing potential blocks on revenue streams. It can also enable cost savings by allowing maintenance interventions to be scheduled only when necessary.
Machine learning techniques provide an ideal solution for solving anomaly detection problems. However, they are typically time-consuming and costly to implement. TIM provides a revolutionary solution to this problem by allowing the development of rigorous anomaly detection models with minimal lead time. This is due to its highly automated and exceptionally fast modeling algorithm.
Due to its speed, TIM’s anomaly detection can easily be applied at scale, to huge numbers of IoT instruments. In addition, the TIM algorithm is extremely lightweight, and can thus be run directly on edge devices, reducing the need for costly network communication.
Finally, TIM’s API-first infrastructure makes it simple to integrate models into a production workflow.
Anomaly detection can be performed on a single instrument output data field, or it can combine information from multiple fields. For example, the information from a number of manufacturing instruments might be used to predict a quality metric for a material being produced. Or multiple data points from a single instrument might be used to predict when failure is likely to occur.
All efforts in marketing and advertising today rely on a wide array of data sources, usually including both internal and external datasets. While this allows for deep insights and highly efficient marketing campaigns, it can also cause problems.
Imagine you are analyzing an ad campaign, when you realize the number of impressions being delivered per day dropped dramatically at a certain date, two weeks ago. A frantic investigation reveals that something has changed in the external data source being used to target potential customers, but the vendor had not alerted you to this. This could equally affect a customer’s insights or segmentation project.
Using machine learning techniques for anomaly detection, you could have detected this ahead of time, instead of discovering the problem weeks or months down the line. However, implementing such a system from scratch requires much time and specialized expertise.
TIM provides a much-needed new approach to this problem, by making it possible to implement robust anomaly detection routines with minimal lead time. Firstly, it is highly automated, meaning no data science experience is required to build effective models. Secondly, the model training process is stunningly fast, taking only a few seconds for a typical dataset – this makes it very easy to build effective models and also allows for huge scaling possibilities. Finally, the API-first infrastructure makes it simple to integrate models into a production workflow.
TIM’s anomaly detection capabilities rest upon first defining “normal behavior” for a given variable or data field (achieved using the TIM forecasting model) and then extending that with an “anomalous behavior” learning algorithm.
Ultimately, an anomaly detection platform with TIM at the center can provide organizations with much-needed confidence in the data that is fundamental to their operation.
TIM’s anomaly detection capabilities can be exploited using a single input variable – for example, if you want to detect anomalies for 1,000 fields from an external data source, you could build one model for each field. It can also be achieved when using multiple input variables. For example, you might want to detect anomalies in conversions, using inputs such as numbers of impressions across multiple marketing channels, economic metrics and more. TIM can handle both types of anomaly detection problems smoothly.
Many cities around the world have adopted a bicycle sharing system. These systems allow people to cycle around the city, picking up a bicycle at one docking station and dropping it off at another. Not all docking stations experience the same use, however. Some stations might be popular starting places, leaving them empty at the end of the day, whereas others see many people arrive throughout the day and end up completely occupied. Understanding and predicting these patterns goes a long way in the planning of the redistribution of bicycles accross docking stations that often happens at night. Yet the situation is even more complex, since people’s behaviour is influenced by many other factors, such as the weather, time of day and calendar. Forecasting the usage of bicycles also gives insight into the demand, allowing cities to anticipate higher demand by expanding the bicycle network, for example.
TIM can help in these scenarios, by taking into account many different variables, transforming and weighing them in order to produce accurate and understandable forecasts. Moreover, TIM allows for frequent model rebuilding, quickly adapting to meet challenges posed by sudden and unexpected changes in peoples behaviour. In case of any issues with data collection, TIM is also able to account for changes in data availability. Thanks to the explainability of the models, it is possible to develop a sense of understanding at what constitutes an accurate forecast, and what influence changes in predictors may have. This can serve as a start for answering questions like “How would the necessary redistribution change if it was colder than expected, tomorrow?”
Apart from historical values of the amount of bicycles in use, this use case takes calendar data, weather data (temperature, windspeed, wind feel, humidity…) and the time of day as input. The generated models then produce forecasted values of the amount of bicycles in use, in the same granularity as the target input data.
Read Success Story