Data analytics: what it is, types, methods and applications

Data analytics is the practice of examining data to answer questions, test explanations and support decisions. It combines preparation, exploration, statistical reasoning and, where appropriate, machine learning. Its value depends on whether the result changes a decision in a useful, measurable way.

A sales forecast can inform purchasing. An analysis of delivery delays can identify where an operation needs attention. Text analytics can reveal recurring issues in service requests. Each requires a different method, an appropriate level of evidence and someone responsible for acting on the findings.

This guide explains the main types of data analytics, the process from question to action, the role of predictive models and the checks needed before analytical results influence operations.

What does data analytics cover?

Data analytics connects a business question with an evidence-based answer. The answer may be a comparison, an explanation worth investigating, a forecast or a recommendation subject to constraints. A chart or a model is an intermediate output; the decision and its consequences determine whether the work was useful.

Analysis can be simple. Comparing repeat purchases across customer groups may require well-defined measures and a spreadsheet rather than a complex model. Other questions involve large datasets, changing behaviour or interactions that justify more specialised methods. Bismart's introduction to data analytics in business explores how analytical work relates to everyday decisions.

Data analytics, business intelligence and data science

Business intelligence organises information for consistent reporting, monitoring and exploration. Data analytics investigates questions and evaluates patterns or possible actions. Data science brings together statistical, computational and domain expertise to develop analytical methods and data products. The boundaries overlap: a BI platform can support exploratory analysis, and an analytics team may build predictive models.

The distinction is useful for assigning work, rather than establishing a hierarchy between disciplines. The comparison of data science and data analytics examines their responsibilities, while the business intelligence guide covers the reporting and information environment that supports recurring decisions.

An analytics programme also needs a clear enterprise data strategy and an explicit definition of data quality for the intended use. A technically successful analysis can still fail if it answers the wrong question or relies on inconsistent records.

What are the main types of data analytics?

Descriptive, diagnostic, predictive and prescriptive analytics describe different questions. They often work together, and a project can move between them as new information becomes available.

  • Descriptive analytics: what happened? Summarises observations, such as weekly sales, average delivery time or the percentage of orders returned.
  • Diagnostic analytics: what might explain it? Examines differences, relationships and process changes. A rise in returns concentrated in one product batch suggests an investigation; it does not by itself establish the cause.
  • Predictive analytics: what is likely to happen? Estimates an unknown outcome, such as next month's demand or the probability of a delayed delivery, with uncertainty.
  • Prescriptive analytics: what action is appropriate? Compares possible actions against objectives and constraints, such as service levels, stock availability or staffing capacity.

A retailer might describe stockouts, investigate their concentration by store, forecast demand and then compare replenishment options. The analysis becomes more demanding as the decision requires stronger assumptions, more timely data or a larger operational commitment. The article on diagnostic analytics develops the investigative part of this sequence.

Where does real-time analytics fit?

Real-time analytics concerns how quickly data is processed and made available. It is not a fifth analytical question. Monitoring a payment stream can require rapid detection, while a quarterly product review may work well with a scheduled refresh. The required latency should follow the decision window and the cost of acting late.

The overview of real-time data analytics in business introduces these applications. Streaming technology does not compensate for unclear definitions, poor data quality or a process that cannot respond to the alert.

The data analytics process: six connected steps

A repeatable process makes results easier to challenge and reuse. The steps below are iterative: a problem discovered during validation can require a return to the original question or the underlying data. The detailed guide to the steps of data analysis provides further background.

  1. Decision and scope. The starting point identifies what can change, who can authorise the change and when a result is needed. It specifies the population, period and scope. “Reduce avoidable delivery delays on domestic orders” is more actionable than “understand logistics”.
  2. Measures and suitable data. Shared definitions establish what counts as a delay, which timestamps are reliable and which records are available at the moment of decision. Permitted access and the necessary level of detail are explicit.
  3. Preparation and exploration. Checks cover missing values, duplicates, inconsistent units, unusual observations and changes in collection methods. Comparing distributions across relevant groups helps inform the choice of model.
  4. Method selection and application. The technique follows the question and the evidence. An understandable baseline provides a reference; additional complexity needs to produce a useful improvement.
  5. Validation and interpretation. Evaluation uses appropriate unseen observations, assesses uncertainty and examines differences between groups or periods. The findings distinguish what the data shows from what remains an assumption.
  6. Action, measurement and review. A named owner takes responsibility for the decision and its rationale. Actual outcomes are compared with expectations. Recurring analytical services also require monitoring of the data and method.

A useful handover contains the question, definitions, method, limitations and proposed action. It also records when the conclusion should be revisited. This is the practical connection between analytical findings and business insights that support decisions.

Organisational adoption matters as much as the analytical technique. The discussion of the gap between business and technology teams addresses that collaboration, and the guide to building a data-driven company considers how evidence becomes part of everyday work.

A starting point for assessing data capabilities

The data maturity model guide helps place analytical work within the capabilities and priorities of the organization.

Advanced analytics, machine learning and algorithms

Advanced analytics usually refers to methods that extend beyond routine reporting, including statistical modelling, machine learning, simulation and optimisation. It does not mean that every problem needs artificial intelligence. The appropriate method is the simplest one that meets the decision's requirements reliably.

The introduction to advanced analytics covers its business scope. An algorithm is a defined procedure for carrying out a task; many algorithms do not learn from data. The related article on algorithms and business problems explains why the objective and operating context must come first.

How machine learning fits into analytics

Machine learning estimates patterns from examples rather than specifying every rule manually. Training adjusts a model using data; evaluation tests whether the resulting behaviour is useful on observations that did not determine those adjustments. Successful training does not establish that the model is ready for operational use.

The explanation of how machine learning algorithms work introduces this mechanism. Planning a project also requires a usable target, suitable data, a baseline and a deployment context, as explored in how to approach a machine learning project.

  • Supervised learning uses examples with target outcomes, such as historical deliveries labelled on time or late.
  • Unsupervised learning explores structure without those target labels, for example by grouping customers with similar purchasing patterns. The groups still require human interpretation.
  • Reinforcement learning learns a policy through interaction and a reward signal. Its suitability depends on the environment, the quality of that signal and the consequences of exploration.

The comparison of supervised and unsupervised learning examines the first two approaches. The broader overview of types of machine learning and their applications provides context for choosing between learning settings.

Classification, regression and clustering

Classification predicts a category or a probability associated with a category. Regression estimates a numerical outcome. Clustering groups observations according to a chosen representation and similarity measure. A cluster is not automatically a meaningful customer segment, and a predicted category is not an established fact.

The practical comparison of classification and clustering develops this distinction. The overview of commonly used machine learning algorithms introduces algorithm families; selection still requires comparison against the task, data and baseline.

When is deep learning appropriate?

Deep learning uses neural networks with multiple processing layers. It can be useful for complex patterns in images, audio, text and other data, but the model's size is not evidence of business value. Available data, pretrained models, infrastructure, latency and oversight all affect suitability.

The guide to machine learning versus deep learning explains their relationship. The article on types of deep learning models introduces different architectures without making them a universal progression that every company must follow.

Preparing data for AI-related analytics

The guide supports a review of sources, preparation, integration and governance before extending the use of data in AI projects.

Predictive analytics and forecasting

Predictive analytics estimates an unknown outcome. A forecast is a prediction tied to a future time or period. Both require a clear target, a decision horizon and a realistic view of which information will be available when the prediction is made.

Classification and regression are common predictive tasks, explored in types of predictive analytics. Time-series forecasting adds concerns such as seasonality, dependence across observations and changes in demand. Randomly mixing past and future observations can make a forecasting evaluation unrealistically favourable.

A useful forecast comparison includes a simple baseline, a test period that respects time order and errors measured at the horizon the business actually uses. The forecasting methods and evaluation guidance in Forecasting: Principles and Practice explains benchmark methods and the use of test data.

Uncertainty and the cost of being wrong

A single estimate can hide material uncertainty. Purchasing decisions may need a plausible range of demand and an assessment of the different costs of excess stock and shortages. A model with the lowest average error may still perform poorly during peak periods or on the products that matter most.

Forecasts support decisions; they do not remove judgement or make the future certain. The discussion of hotel occupancy forecasting with machine learning illustrates a setting in which lead time, seasonality and the operating decision need to be considered together.

Simulation and digital twins

Simulation compares outcomes under stated assumptions. A digital twin represents a real entity or process and connects that representation with information about its state. Its usefulness depends on the fidelity of the representation and its validation for the intended decision.

A dashboard or forecast does not automatically constitute a digital twin. Bismart's introduction to digital twins explores this approach. Simulated improvements remain estimates until their assumptions and operational effects are tested.

A checklist for reviewing actionable insights

The template connects findings, hypotheses, validation and possible actions before an analytical conclusion is used.

Data mining and text analytics

Data mining searches for patterns and relationships in datasets. The patterns may suggest useful hypotheses, but their discovery does not establish causation or guarantee that they will remain useful on new observations. The guide to data mining in business introduces its role in analysis.

Text analytics applies analytical methods to written language, such as support tickets, research documents or comments. Depending on the objective, it can identify entities, assign categories, group topics or extract recurring terms. The overview of text analytics and practical examples develops these applications.

From documents to usable evidence

A text analytics project needs a clear unit of analysis: a document, a passage, a ticket or an individual mention. Evaluation should consider language, terminology, missing context and the consequences of incorrect extraction. A fluent summary alone is not proof that every statement is supported by the source documents.

Taxonomies organise information using a defined structure; folksonomies traditionally emerge from users' tags. Automated tagging and Bismart's Folksonomy solution are related applications, rather than interchangeable definitions. The comparison of folksonomy and taxonomy explains the distinction, with examples of folksonomy and an introduction to intelligent folksonomy solutions providing further context.

Search and retrieval are another part of this workflow. The article on Elasticsearch projects at Bismart describes an implementation context. Speech-based inputs add a transcription stage whose errors can affect subsequent analysis, as discussed in the introduction to speech recognition applications.

A documented healthcare research application

The Hospital del Mar and Grupo Ferrer project describes the use of text analytics in a clinical research setting. The associated Folksonomy Text Analytics webinar presents the project and its analytical approach. These are examples of a specific collaboration, rather than evidence that the same outcomes can be assumed in another organisation or that analytical outputs replace clinical validation.

How should analytical results be validated?

Validation asks whether a result is credible and useful for the intended decision. The required evidence depends on the consequences of error, the data and the operating conditions. A model can perform well on an average metric and still be unsuitable for a particular use.

Use appropriate metrics and prevent data leakage

For numerical predictions, measures such as mean absolute error express the typical size of an error in the target's units. For classification, precision and recall describe different aspects of performance. When important cases are rare, overall accuracy can be misleading. The scikit-learn model evaluation documentation explains these measures and their limitations.

Data leakage occurs when information that should be unavailable during training or prediction influences the model or its evaluation. Examples include preparing data using the entire dataset before splitting it, or predicting a delay using a field only completed after delivery. The scikit-learn guidance on data leakage explains why preprocessing must respect the separation between training and evaluation data.

Distinguish association from causation

A relationship between a promotion and higher sales does not establish how much of the increase the promotion caused. Seasonality, customer selection or other changes may explain part of the difference. Where appropriate, a planned experiment can provide stronger evidence about an intervention. The NIST introduction to experimental design explains how controlled changes to inputs support conclusions about responses.

Even an experiment needs a defined outcome, suitable comparison and attention to practical limitations. Where randomisation is unavailable, assumptions and alternative explanations should be made explicit rather than hidden behind a causal claim.

Monitor results after deployment

Operational conditions change. Data distributions, missing values, processing times and model performance can move away from their original values. A monitoring plan should specify signals, thresholds, an owner and a response, including when to investigate, retrain, suspend or return to a simpler process.

Input drift is a signal for investigation, not proof that predictions have become wrong. Where outcomes become available, performance should also be measured against them. Microsoft's model monitoring documentation distinguishes several monitoring signals used in operational machine learning.

Software testing is a related engineering activity with a different purpose. The article on machine learning and AI in test automation concerns the testing of software; it does not replace evaluation of a model's analytical validity or business effects.

How should data analytics tools be selected?

Tool selection should follow the decision, required methods and operating model. The evaluation needs to cover data access, analytical capabilities, collaboration, repeatability, permissions, deployment, monitoring and total cost. A familiar interface may help adoption, but it does not remove the need to validate the work.

  • Spreadsheets and SQL can support exploration, reconciliation and repeatable queries when scale and controls are appropriate.
  • BI and visual analysis tools support shared measures and investigation, often alongside statistical or programming environments.
  • Statistical and machine learning environments support model development, testing and reproducible workflows.
  • Text and search platforms support indexing, retrieval and language-based analysis, with evaluation matched to the task.

Bismart's roundups of data analytics tools and data science tools provide starting points for comparing categories. More specialised overviews cover machine learning tools, deep learning tools and text analytics systems. Product capabilities, licensing and availability should be checked against current vendor documentation before a purchase decision.

The historical article on Bismart Folksonomy's inclusion in a 2023 unstructured data management market report provides company and market context. It is not an independent current product comparison or a ranked evaluation.

Analytical outputs can be incorporated into intelligence-enriched business applications or AI-enabled business automation. This introduces additional requirements: controlled access, error handling, human review where needed and a way to measure the consequences of acting on a prediction.

An application in hotel management

The case study provides a sector example of organizing information for hotel management and business decisions. Outcomes belong to the project described.

Business applications: the decision comes first

In logistics, a prediction of late arrival is useful only if there is enough time and authority to change the response. The overview of data analytics for logistics introduces this connection between analysis and operations. In customer service, analysing recurring complaint categories can inform process changes, provided the categories are consistent and representative.

Published sector examples can help frame questions, but they should not be treated as proof that another company will obtain the same result. Bismart's articles on data intelligence in retail and Zara and Amazon and predictive logistics provide background reading. A described patent or proposed operating model is not evidence that it has been deployed at scale.

A historical account of the Bismart and Libelium environmental monitoring project in Brussels provides a further example of combining operational observations with analytical questions. Its project context and time frame matter; it should not be read as a current clinical recommendation.

Worked example: forecasting demand for a service team

Illustrative scenario: a service manager must set next week's staffing schedule. The current reference is demand for the equivalent weekday in recent weeks. The proposed improvement uses historical ticket arrivals and known calendar information.

  1. Target definition. The model forecasts incoming tickets by day and relevant service category, using information available before the schedule is approved.
  2. Baseline comparison. The proposed model is compared with a simple historical reference over the same periods.
  3. Realistic evaluation. Testing uses later periods, including busy weeks, and examines both typical error and underestimation when demand is high.
  4. Translation into staffing. The calculation uses stated assumptions about handling time, availability and service levels. These assumptions require separate validation.
  5. Controlled pilot. Forecasts are reviewed with the manager, overrides are recorded and evaluation covers service performance, overtime and cost.

The result is not merely a better forecast score. It is evidence about whether the staffing process improves without unacceptable effects on service or employees. The rollout decision should record limitations, responsibilities and a review date.

When the next step is an implementation project, Bismart's machine learning and predictive analytics consulting connects forecasting, segmentation and operational use cases with the organisation's data architecture and business processes.

Eight tips for connecting evidence and decisions

The ebook develops practical considerations for using data in a decision process and interpreting the resulting evidence.

Practical resources for moving from analysis to action

The following Bismart resources address different parts of the analytical workflow. They complement the guide; they are not substitutes for validating a specific model or business decision.

Frequently asked questions about data analytics

What is the difference between data analysis and data analytics?

The terms overlap. Data analysis often describes the work of examining a dataset to answer a question; data analytics can also refer to the broader methods, systems and practices that support that work. A useful project defines its question and deliverable rather than relying on a rigid distinction between the terms.

Does data analytics always require machine learning?

No. Comparisons, aggregation, visual exploration and statistical analysis can answer many business questions. Machine learning is appropriate when it offers a useful, validated improvement over a simpler approach and can be maintained within operational constraints.

How much data is needed for predictive analytics?

There is no universal minimum. Requirements depend on the task, variability, number of relevant cases, target quality and method. A large dataset with missing outcomes or misleading selection can be less useful than a smaller dataset that represents the intended decision.

How can the business value of analytics be measured?

Measurement connects the analytical output to a defined decision and compares outcomes against an appropriate reference. Relevant measures may include avoidable cost, service quality, time saved, stock availability or risk exposure. Model performance and business impact should be assessed separately because an improvement in one does not guarantee an improvement in the other.

What should be documented before a model is used in production?

The intended use, data definitions, evaluation method, performance limitations, permitted users and operational owner should be recorded. The documentation should also specify monitoring, response to failure and the conditions for reviewing or withdrawing the model.