Model guides

Browse forecasting, prediction, and classification services by capability

Each model guide includes endpoint details, route structure, and example request and response shapes across time-series forecasting, feature-based prediction, and classification-oriented services.

Time-series forecasting

Forecast sales, demand, product usage, customer activity, revenue, capacity, operations, and service health from historical signal patterns.

Best when recent history carries enough structure to estimate what is likely to happen next.

  • ARIMA for autoregressive forecasting
  • SARIMA for seasonal patterns
  • SARIMAX for seasonal forecasting with exogenous features
  • AutoReg for lightweight temporal inference

Feature-based prediction

Predict counts and numeric outcomes using product, customer, sales, demand, operational, business, and service context.

Useful when the next outcome depends on product usage, demand, customer attributes, campaign activity, inventory, capacity, revenue movement, service health, or other business features.

  • Poisson for count-based outcomes
  • Linear and Polynomial for interpretable baselines and curved feature relationships
  • Tree and Forest paths for non-linear behavior

Classification services

Serve churn risk, purchase propensity, customer state, risk class, and decision-category workflows.

Useful when the goal is to classify an outcome such as customer state, purchase propensity, risk class, or another business decision category.

  • Logistic Regression for interpretable probability-based classification
  • Gradient Boosting Classifier for boosted decision-tree classification across complex feature interactions
  • Naive Bayes for lightweight probabilistic classification

What each guide includes

Open any model below to see the route-level details

Each model page explains endpoint shape, sample requests, response format, and trust-aware interpretation.

Endpoint paths

Find the exact route structure and service path for the selected model.

Payload expectations

Review the input structure, required fields, and example request bodies.

Response and trust behavior

Understand the response shape, prediction output, and how trust-aware fields should be interpreted.

How to choose a model

Pick the forecasting, prediction, or classification service that matches the signal and the decision you need to make

The right starting point depends on whether history alone explains the next outcome, whether additional product, customer, sales, demand, operational, business, or service context needs to be part of the prediction, or whether the goal is to classify a business or customer outcome.

Use time-series models

Start here when past values already contain useful forecasting structure.

  • Best for historical numeric signals over time
  • Useful when trend or seasonality matters
  • Good for baseline forecasting from recent signal history

Use feature-based models

Start here when product, customer, sales, demand, operational, business, or service context changes the outcome.

  • Best when contextual signals matter
  • Useful for count or numeric prediction with product, customer, sales, operational, business, or service features
  • Helps when history alone is not enough

Use classification services

Start here when the goal is to classify a business or customer outcome.

  • Best for purchase propensity, behavioral scoring, or decision categories
  • Useful when the response is a class, label, or probability-based outcome
  • Supports API-driven customer and business decision workflows

Start with a simple baseline

Choose the lightest useful model before moving into more complex paths.

  • Use Linear or AutoReg as practical baselines
  • Measure whether extra complexity actually improves the result
  • Keep interpretability high during early validation

Match the output to the decision

The model should fit not only the data, but also the action the response will drive.

  • Use forecasting when timing and trend matter most
  • Use prediction when live feature context changes the outcome
  • Use classification when the result is a label, probability, or decision category

Compare outputs for trust

Use more than one path when prediction quality matters more than convenience.

  • Compare baseline and expressive models
  • Check disagreement before acting
  • Use trust warnings and suppression policies in production

Supported models

Open a model to view endpoints, requests, and responses

These cards are meant to help you choose where to start, not just browse names. Each model has a different role depending on the signal shape, feature context, and trust needs.

Linear Regression

A simple and interpretable baseline model that predicts a continuous target as a linear combination of input features. Useful when you want to understand how product, customer, sales, demand, revenue, operational, or service-health variables relate to the predicted outcome. Fast to train, easy to explain, and often a good first comparison point before moving to more complex models.

Best for: interpretable baseline prediction from live features

Input type: feature request with product, customer, sales, operational, business, or service context

Use when: you want a simple and explainable feature-to-output relationship

predict/linear

Polynomial Regression

A regression model that extends linear regression by adding polynomial feature terms such as x², x³, or higher-order combinations. Useful when the relationship between business, product, customer, sales, demand, operational, or service inputs and the predicted outcome is curved rather than purely straight-line. A good choice when you want more flexibility than a simple linear baseline while still staying relatively interpretable compared to more complex tree-based models.

Best for: curved feature-to-output relationships that need more flexibility than a straight-line baseline

Input type: feature request with product, customer, sales, operational, business, or service context

Use when: the route fits the signal shape and production decision context

predict/polynomial

Poisson Regression

A generalized linear model designed for count-style targets such as units sold, orders, transactions, usage counts, event frequency, or service-health totals over a time window. Useful when the prediction target is a non-negative integer and you want a model that fits count behavior more naturally than ordinary linear regression. A strong choice when counts matter more than continuous values.

Best for: count-style outcomes such as units sold, usage counts, transaction counts, capacity totals, event totals, or service-health totals

Input type: feature request with product, customer, sales, operational, business, or service context

Use when: the prediction target behaves like counts over a window

predict/poisson

Decision Tree Regressor

A non-linear tree-based model that partitions the feature space into decision regions based on learned thresholds. Useful when business, product, customer, demand, or service behavior changes in step-like ways, such as after demand crosses a limit or a customer segment enters a different state. Good for capturing interactions and non-linear patterns with decision logic that is often easier to inspect than more complex ensembles.

Best for: threshold-driven non-linear patterns with simple decision paths

Input type: feature request with product, customer, sales, operational, business, or service context

Use when: clear rule splits matter more than smooth mathematical relationships

predict/decision_tree

Random Forest Regressor

An ensemble of decision trees that improves robustness by averaging predictions across many trees instead of relying on a single path. Useful when outcomes depend on non-linear feature interactions across product usage, customer behavior, sales activity, demand, revenue, operations, or service health. Often a strong default when you want more flexibility than a linear model and more stability than a single decision tree.

Best for: richer non-linear feature interactions across product, customer, sales, operational, business, or service context

Input type: feature request with product, customer, sales, operational, business, or service context

Use when: feature interactions are too complex for a single simple model

predict/randomforest

ARIMA

Autoregressive Integrated Moving Average, a classical time-series model that forecasts future values from past observations using autoregressive and moving-average structure with differencing for trend handling. Useful when recent historical behavior carries meaningful signal into the near future and you want a proven baseline for univariate forecasting across sales, demand, usage, revenue, capacity, or service-health signals.

Best for: historical numeric forecasting from recent past behavior

Input type: time-series signal history

Use when: the next value depends mainly on recent historical behavior

timeseries/arima

SARIMA

Seasonal ARIMA, an extension of ARIMA that adds seasonal differencing and seasonal autoregressive and moving-average structure. Useful for sales, demand, usage, revenue, capacity, or service-health signals that repeat in hourly, daily, weekly, or other regular cycles. A strong choice when the metric has recurring seasonal behavior and you want the forecast to reflect both local trend and repeating patterns.

Best for: sales, demand, usage, revenue, capacity, or service-health signals with repeating cycles

Input type: time-series signal history

Use when: the signal shows meaningful seasonality across repeated time windows

timeseries/sarima

AutoARIMA

An automated ARIMA selection approach that searches across candidate configurations and uses information criteria to find a reasonable time-series fit. Useful when you want the convenience of ARIMA-style forecasting without manually tuning the order parameters first. A good starting point for teams that want a practical automatic baseline before hand-selecting a more specific time-series setup.

Best for: historical numeric forecasting from recent past behavior

Input type: time-series signal history

Use when: the next value depends mainly on recent historical behavior

timeseries/autoarima

AutoReg (Autoregression)

An autoregression model that uses lagged historical values as direct predictors for the next value in the series. Useful for short-memory sales, demand, usage, revenue, capacity, or service-health signals where recent past behavior strongly influences the near future. Simple, fast, and interpretable, making it a good lightweight baseline before moving to more expressive seasonal or feature-based approaches.

Best for: simple baseline forecasting from recent signal history

Input type: time-series signal history

Use when: you want a lightweight baseline before trying more complex paths

timeseries/autoregression

Trust-aware behavior

Some predictions should be explained carefully. Some should be suppressed.

eazePredict does not treat every model output as equally safe. Depending on the input regime, the platform may return warnings, expose disagreement, or suppress the numeric prediction entirely.

Trust warnings

Responses may include trust_warning when the system detects unstable or suspicious operating conditions.

Prediction suppression

In extreme cases, numeric prediction values may be returned as null instead of pretending the output is reliable.

Disagreement visibility

Model spread and regime checks help teams understand when the forecast surface is behaving unpredictably.