Credit and risk scoring across consumer finance, real estate, insurance & reinsurance and financial markets are commonly done with linear models due to their reliability and interpretability. Whenever these models need to be integrated into an app, website, or internal service they need to be deployed to the web, usually as an API.

Sales and demand forecasts are commonly done via linear models, as these often have good performance and high levels of reliability. Other common applications include marketing ROI modelling, pricing optimization and inventory management. If these need to be regularly done based on changing data, and done consistently across time and the organisation, deployment via API is your friend.

Customer segment classification, decision modelling and quantitative market modelling are generally done via linear models. Deploying these via API expands their effectiveness for customers.

A two-step tutorial shows how to create a customer segmentation and how to identify segment membership in production. Deploying a segment classifier on diagonal.sh enables questionnaire designs with different questionnaire items for different segments. For each respondent, a simple call to the model API establishes which segment they belong to, so that the subsequent questions can be tailored to that segment.

Forecasting models, risk assessment models, macroeconomic models and many other kinds of models routinely produced in economic and policy consulting benefit from API deployment, which enables customers to evaluate their situation using new data as it becomes available

Customer churn prediction, newsletter optimization, UI personalization and many other modelling tasks can be done using linear models. Once deployed via API, they can improve the user experience and increase company effectiveness.

Anomaly detection, threat scoring and fraud detection are common use cases for linear models. Through API deployment, these techniques can be applied consistently through an organization or product. diagonal.sh enables frequent updating of these models as the threat landscape evolves.

Content optimization and gameplay adaptation can be dynamically learned from user behavior using linear models, and quickly fed back into the game play experience via APIs.

Impact modelling and forecasting of more general (macroeconomic or ecological, for example) and more specific variables in international development are frequently done using linear models, as they guarantee reliability and interpretability. Deploying them via API can make these models useful and responsive to new data across the world.

Linear regression models enable the prediction of a real-valued variable from a single or multiple categorical or real-valued variables.

For example, if we want to predict the crop yield of a field based on the weather, the viability of a new outlet in a location given its geography, or the price of a security from financial variables, each of these relationships can be modelled using linear regression

Linear regression has some variants from the most common (ordinary least squares) type, such as ridge regression and lasso regression. These variants can also be exported to diagonal.

Logistic regression enables the prediction of a categorical variable and their corresponding probabilities from a single or multiple categorical or real-valued variables.

For example, if we want to predict the risk of wildfire in a region based on climate, weather and vegetation, the susceptibility of crops to fall victim to a specific disease, or the risk of default of a borrower, these relationships with binary or categorical outcomes can be modelled with a logistic regression.

Generally, whenever one of several discrete actions can be taken or not taken, a discrete representation of the model output space becomes necessary. With logistic regression, that discreteness is incorporated into the model itself.

The ridge classifier variant of logistic regression can also be deployed to diagonal.

In our tutorial on segment classification, we show one of the many applications of logistic regression deployments.

You can update a model that is deployed on diagonal via console or API call.

This enables a form of online learning where a model is continuously updated as new data comes in. The delay between model update and deployment means that immediate adaptation is not possible, but with a delay of about 10 to 15 minutes. For many applications, this is already very useful.

Online learning is useful whenever the relationship between the input variables and the output variable fluctuates or changes over time. Through model updates or retraining, the most recent relationship is discovered and applied. These models are much less common than linear regression or logistic regression, but often find application in finance, forecasting and industrial applications.

Contextual multi-armed bandits are a class of algorithms that learn to select the best action out of a predetermined set of actions, based on context.

For example, imagine you run some LLM-based chat service, and run into capacity constraints at peak times. Different users might be more annoyed by slow responses, and other by a notice that their query cannot be processed currently (this option has the advantage of being immediate). A contextual bandit would allow you to select which of these two usage constraints to apply to a given user, so as to maximize customer retention.

Contextual multi-armed bandits can be deployed on diagonal by deploying several linear models and updating them regularly, i.e. multiple online learning models.