AI Use Cases: Forecasting and Anomaly Detection

AI Scenario Selection: Where to Start?

An AI project’s fate is usually sealed by one early decision — picking the right scenario — long before a line of code is written

The nature of the business problem largely dictates which model architecture and data structure you will need, and that is exactly why AI use cases differ so much from one another. Choose the wrong scenario and you end up with projects that are technically “working” yet deliver no business value at all.

Before you commit to a scenario, ask yourself the following questions:

• What is the output? A number (regression), a category (classification), or a normal/abnormal split (anomaly)?
• What is the data structure? Tabular, text, image, or time series?
• Is there labeled data? Supervised or unsupervised?
• Real-time or batch? An instant decision or a periodic analysis?

5 Core AI Scenario Categories

In practice, operational AI applications fall into five main categories:

1. Forecasting (Regression): Predicting future numerical values
2. Classification: Categorizing input into predefined classes
3. Anomaly Detection: Identifying non-normal data points
4. Natural Language Processing (NLP): Extracting meaning from text data
5. Computer Vision: Extracting information from visual data

Forecasting Models

Forecasting models read the future from past data, so you can act before an event rather than after it

Forecasting models learn from historical data to predict future numerical values. It is one of the most common AI use cases you will run into in the field.

Types of Forecasting Models

Time Series Forecasting

Predicts future periods in time-dependent data:

• ARIMA/SARIMA: Statistical approach with seasonality support
• Prophet: Optimized for holiday effects and trend changes
• LSTM: Complex patterns via deep learning
• XGBoost/LightGBM: Enhanced forecasting with feature engineering

Regression (Numerical Prediction)

Predicts a dependent variable from independent variables:

• Linear Regression: Simple, interpretable, baseline model
• Random Forest Regression: Capturing non-linear relationships
• Gradient Boosting: High accuracy, complex data structures

Forecasting Use Cases

Data Requirements

• Minimum data duration: 12-24 months (2 years recommended to capture seasonality)
• Data frequency: Daily, weekly, or monthly (problem-dependent)
• Additional variables: Promotion calendar, holidays, economic indicators
• Data quality: Missing values should stay below 5%

Success Metrics

• RMSE (Root Mean Square Error): Average error magnitude
• MAE (Mean Absolute Error): Absolute mean deviation
• MAPE (Mean Absolute Percentage Error): Percentage error (target: below 10-15%)
• Forecast Bias: Tendency for systemic over/under-forecasting

Classification

Classification is where you sort data into buckets automatically and let the decisions move faster

Classification models place input into predefined categories (classes). The output is discrete and categorical — A/B/C, Yes/No, Low/Medium/High.

Types of Classification Models

Binary Classification

Distinguishes between two classes:

• Churn prediction: Will stay / Will leave
• Credit risk assessment: Approve / Reject
• Spam detection: Spam / Not Spam

Multi-class Classification

Distinguishes between more than two classes:

• Customer segmentation: A / B / C / D segment
• Product categorization: Automatic assignment to 10+ categories
• Support ticket routing: Technical / Billing / Return / Other

Multi-label Classification

An input can belong to multiple classes:

• Document tagging: A document can be both “Contract” and “Confidential”
• Product features: A product carries multiple attributes

Classification Use Cases

Common Algorithms

• Logistic Regression: Simple, interpretable, baseline
• Random Forest: Balanced performance, resistant to overfitting
• XGBoost/LightGBM: High accuracy, competition winner
• SVM (Support Vector Machine): Effective for high-dimensional data
• Neural Networks: Complex patterns, large datasets

Success Metrics

• Accuracy: Correct prediction rate (for balanced classes)
• Precision: Ratio of positive predictions that are truly positive
• Recall: Ratio of actual positives captured
• F1 Score: Balance between Precision and Recall (target: >0.75)
• AUC-ROC: Model discriminative power

Anomaly Detection

Anomaly detection catches what falls outside “normal” automatically, surfacing risk before it has a chance to grow

Anomaly detection automatically flags the non-normal points (outliers) in your data. The logic is simple: it profiles normal behavior first, then marks anything that does not fit that profile.

Anomaly Detection Approaches

Statistical Methods

• Z-Score: Deviation measurement based on standard deviation
• IQR (Interquartile Range): Outlier detection based on box plots
• Moving Average: Trend deviations in time series

Machine Learning Methods

• Isolation Forest: Isolation-based, fast and scalable
• One-Class SVM: Training with only normal data
• DBSCAN: Outlier detection via density-based clustering
• Autoencoder: Deep learning, complex patterns

Rule-Based Methods

• Business rules: “Invoice amount 10x average is an anomaly”
• Threshold alerts: Predefined threshold values
• Hybrid approach: ML + business rules combined

Anomaly Detection Use Cases

Data Requirements

• Normal data heavy: Anomalies are usually 1-5%
• Labeled data (for supervised): Known anomaly examples
• Timestamp: Critical for temporal anomalies
• Multi-variable: Situations that are normal individually but anomalous together

Success Metrics

• Precision: Rate of anomaly alarms that are truly anomalies
• Recall: Rate of actual anomalies captured
• False Positive Rate: False alarm rate (should be kept low)
• Detection Latency: Delay time in anomaly detection

NLP Applications (Natural Language Processing)

NLP pulls meaning out of messy, unstructured text and turns it into something you can automate against

Natural Language Processing (NLP) lets you extract, analyze, and generate meaning from text and speech data. In short, it turns the unstructured data sitting in your business workflows into something processable.

NLP Task Types

Text Classification

• Sentiment analysis: Positive / Negative / Neutral
• Topic classification: Email, support ticket routing
• Spam/ham detection
• Intent classification

Information Extraction

• Named Entity Recognition (NER): Person, company, date, amount
• Relationship extraction: Connections between entities
• Keyword extraction
• Summarization

Text Generation

• Automatic response suggestions
• Drafting reports
• Translations

NLP Use Cases

NLP Technical Infrastructure

• Tokenization: Splitting text into words/sub-words
• Word Embeddings: Word2Vec, GloVe, FastText
• Transformer Models: BERT, GPT, T5 family
• Pre-trained Models: BERTurk, mBERT for Turkish

Challenges

• Turkish NLP: Agglutinative language structure, resource limitations
• Domain-specific vocabulary: Adaptation to industry jargon
• Ambiguity: Different meanings of the same word
• Data privacy: Texts containing personal data

Computer Vision

Computer vision extracts information from images, moving inspections that were once done by eye into the digital world

Computer Vision analyzes visual data (photos, video) to extract information from it. It is widely used in production, logistics, and quality control.

Computer Vision Task Types

Image Classification

• Determining which category an image belongs to
• Example: Is the product defective or flawless?

Object Detection

• Detecting objects in an image along with their locations
• Example: How many products are on the shelf, where are they?
• Algorithms: YOLO, Faster R-CNN, SSD

Segmentation

• Classifying every pixel in an image
• Semantic segmentation: Class-based
• Instance segmentation: Object-based distinction

OCR (Optical Character Recognition)

• Digitizing text content in images
• Reading invoices, labels, license plates

Operational Use Cases

Technical Infrastructure

• CNN (Convolutional Neural Networks): Basic image processing architecture
• Transfer Learning: Pre-trained models (ResNet, VGG, EfficientNet)
• Edge Computing: Processing near the camera (low latency)
• GPU Infrastructure: Required for training and inference

Data Requirements

• Image quality: Sufficient resolution and lighting
• Labeled data: 100-1000+ images per class
• Diversity: Different angles, lighting, background conditions
• Balance: Balance of sample counts between classes

Scenario Comparison Table

Review the comparison table below to see which AI scenario fits which situation:

Scenario Selection Flow

Follow these steps to select the right scenario:

1. Define the business problem: “Which decision do I want to automate?”
2. Determine the output type: Number, category, or normal/abnormal?
3. Evaluate the data structure: Tabular, text, image, or time series?
4. Check the status of labeled data: Supervised or unsupervised?
5. Select the scenario based on the table: Refer to the table above

Frequently Asked Questions (FAQ)