Glossary

ARIMA

A classical statistical model for analyzing and forecasting time-series data that leverages autocorrelation, differencing, and moving averages to capture linear patterns in lead time history.

Get in touch Learn more

Data scientist building training data pipeline on laptop, data preprocessing visible, technical workspace.

TIME-SERIES FORECASTING

What is ARIMA?

ARIMA is a classical statistical model for analyzing and forecasting time-series data by leveraging autocorrelation, differencing, and moving averages to capture linear patterns.

An Autoregressive Integrated Moving Average (ARIMA) model is a statistical analysis tool that uses past values and errors to forecast future points in a time series. It decomposes a signal into three components: autoregression (AR) , which models the dependency between an observation and lagged observations; integration (I) , which applies differencing to make the data stationary; and a moving average (MA) , which models the dependency between an observation and residual errors from a moving average model applied to lagged observations.

ARIMA is explicitly designed for univariate data exhibiting stable, linear trends and is specified by the notation ARIMA(p, d, q). The p parameter defines the lag order, d is the degree of differencing required to remove trends, and q defines the order of the moving average window. While highly effective for stable demand patterns, ARIMA struggles with non-linear volatility and requires manual intervention to handle concept drift, often serving as a baseline before deploying more complex models like LSTM or Temporal Fusion Transformers.

CORE PROPERTIES

Key Characteristics of ARIMA Models

ARIMA models are defined by three core components—autoregression, integration, and moving averages—that work together to capture linear patterns in stationary time-series data.

Autoregressive (AR) Component

The AR term models the dependency between an observation and a specified number of lagged observations (p). It captures the momentum or inertia in a time series.

Mechanism: Regresses the variable on its own prior values
Parameter p: The number of lag observations included in the model
Example: A lead time series where today's delay is correlated with yesterday's delay uses AR(1)
Diagnostic: Identified by decaying spikes in the Partial Autocorrelation Function (PACF) plot

Integrated (I) Differencing

The I component applies differencing to make a non-stationary time series stationary by removing trend and seasonality. This is a critical preprocessing step before modeling.

Order d: The number of times the data is differenced
First-order differencing: Subtracts the previous observation from the current one
Stationarity requirement: ARIMA assumes constant mean, variance, and autocorrelation over time
Over-differencing risk: Excessive differencing introduces unnecessary noise and reduces forecast accuracy

Moving Average (MA) Component

The MA term models the dependency between an observation and the residual errors from a moving average applied to lagged forecast errors (q). It captures shock effects.

Mechanism: Uses past forecast errors to correct future predictions
Parameter q: The size of the moving average window of error terms
Example: A one-time port closure creates a shock that propagates through subsequent lead time forecasts
Diagnostic: Identified by a sharp cutoff in the Autocorrelation Function (ACF) plot after lag q

Seasonal ARIMA (SARIMA)

SARIMA extends the standard ARIMA framework by adding seasonal autoregressive, differencing, and moving average terms to capture repeating cyclical patterns.

Notation: SARIMA(p,d,q)(P,D,Q)m where m is the seasonal period
Seasonal P,D,Q: Separate parameters for the seasonal component
Example: Monthly lead time data with annual seasonality uses m=12
Application: Critical for supply chains with predictable seasonal demand spikes like holiday retail or agricultural harvest cycles

Box-Jenkins Methodology

The systematic three-stage approach for building ARIMA models, developed by statisticians George Box and Gwilym Jenkins in 1970.

Stage 1 - Identification: Use ACF and PACF plots to determine initial p, d, q values
Stage 2 - Estimation: Fit the model parameters using maximum likelihood estimation
Stage 3 - Diagnostic Checking: Analyze residuals to confirm they resemble white noise with no remaining autocorrelation
Iterative process: The cycle repeats until residuals are statistically indistinguishable from random noise

Stationarity and the Augmented Dickey-Fuller Test

The Augmented Dickey-Fuller (ADF) test is the standard statistical hypothesis test used to verify that a differenced series has achieved stationarity before fitting an ARIMA model.

Null hypothesis: The time series has a unit root and is non-stationary
Interpretation: A p-value below 0.05 rejects the null, confirming stationarity
Practical use: Determines the required differencing order d
Limitation: ARIMA cannot model structural breaks or regime changes without manual intervention

MODEL SELECTION GUIDE

ARIMA vs. Modern Forecasting Alternatives

Comparative analysis of ARIMA against contemporary machine learning approaches for lead time forecasting, highlighting key architectural differences and deployment considerations.

Feature	ARIMA	Gradient Boosting (GBM)	Temporal Fusion Transformer (TFT)
Model Family	Classical Statistical	Ensemble Tree-Based	Attention-Based Deep Learning
Handles Non-Linear Patterns
Native Uncertainty Quantification
Exogenous Variable Support
Multi-Horizon Forecasting
Interpretability	High (coefficients)	Medium (feature importance)	High (attention weights)
Cold-Start Performance	Poor (needs 50+ obs)	Moderate	Poor (needs large data)
Training Speed	< 1 sec	Seconds to minutes	Minutes to hours

ARIMA FORECASTING

Frequently Asked Questions

Clear, technical answers to the most common questions about ARIMA modeling for supply chain time-series analysis.

ARIMA (Autoregressive Integrated Moving Average) is a classical statistical model for analyzing and forecasting univariate time-series data. It decomposes a time series into three components: Autoregression (AR), which models the dependency between an observation and a specified number of lagged observations; Integration (I), which applies differencing to make the series stationary by removing trends; and Moving Average (MA), which models the dependency between an observation and the residual errors from a moving average applied to lagged observations. The model is denoted as ARIMA(p,d,q), where p is the order of the AR term, d is the degree of differencing, and q is the order of the MA term. For lead time forecasting, ARIMA captures linear patterns in historical supplier delivery data to project future performance.

Enabling Efficiency, Speed & Accuracy

Intelligent Analysis, Decision & Execution

We build AI systems for teams that need search across company data, workflow automation across tools, or AI features inside products and internal software.

Talk to Us

Search across company data

Give teams answers from docs, tickets, runbooks, and product data with sources and permissions.

Useful when people spend too long searching or get different answers from different systems.

Enterprise searchRAGPermissions

Automate internal workflows

Use AI to route work, draft outputs, trigger actions, and keep approvals and logs in place.

Useful when repetitive work moves across multiple tools and teams.

AI agentsWorkflow automationGovernance

Add AI to products and internal tools

Build assistants, guided actions, or decision support into the software your team or customers already use.

Useful when AI needs to be part of the product, not a separate tool.

AI integrationDecision supportModel routing

TIME-SERIES FORECASTING ECOSYSTEM

Related Terms

ARIMA is a foundational model for lead time analytics. These related concepts define the broader statistical and machine learning landscape for predicting delivery performance.

Probabilistic Forecasting

Unlike ARIMA's point forecasts, probabilistic methods output a distribution of possible outcomes with quantified uncertainty. This is critical for supply chains where a single deterministic lead time estimate is insufficient for risk management.

Generates prediction intervals (e.g., 80% confidence)
Enables safety stock calculations based on service level targets
Often implemented via Quantile Regression or Bayesian methods

80-95%

Typical Confidence Interval

Long Short-Term Memory (LSTM)

A specialized recurrent neural network that captures long-term dependencies in sequential data. While ARIMA excels at linear patterns, LSTMs model complex non-linear relationships in lead time history.

Remembers patterns across hundreds of time steps
Handles multivariate inputs (weather, port congestion, seasonality)
Requires significantly more training data than ARIMA

Concept Drift

The phenomenon where the statistical properties of lead time change over time, violating ARIMA's stationarity assumption. A model trained on pre-pandemic data will fail when supplier behavior fundamentally shifts.

Sudden drift: Abrupt change from a port closure or tariff
Gradual drift: Slow degradation of supplier performance
Requires continuous Model Drift Monitoring to trigger retraining

Prediction Intervals

A range around ARIMA's point forecast where future lead times are expected to fall with a specified probability. Critical for translating forecasts into actionable buffer stock decisions.

ARIMA generates intervals assuming normally distributed errors
Conformal Prediction provides distribution-free guarantees
Wider intervals signal higher supply chain risk exposure

Gradient Boosting Machine (GBM)

An ensemble method that sequentially builds decision trees to correct prior errors. GBMs often outperform ARIMA on tabular lead time data with rich features like carrier type, origin port, and season.

Handles mixed data types (categorical + numerical)
Automatically captures non-linear interactions
Popular implementations: XGBoost, LightGBM, CatBoost

Temporal Fusion Transformer (TFT)

A state-of-the-art attention-based model for interpretable multi-horizon forecasting. TFT explicitly handles static covariates (supplier category) and known future inputs (planned holidays) that ARIMA cannot natively incorporate.

Variable selection networks identify key drivers
Quantile outputs for uncertainty quantification
Outperforms ARIMA on complex, multi-seasonal supply chain data

About the author

Prasad Kumkar

CEO & MD, Inference Systems

Prasad Kumkar is the CEO & MD of Inference Systems and writes about AI systems architecture, LLM infrastructure, model serving, evaluation, and production deployment. Over 5+ years, he has worked across computer vision models, L5 autonomous vehicle systems, and LLM research, with a focus on taking complex AI ideas into real-world engineering systems.

His work and writing cover AI systems, large language models, AI agents, multimodal systems, autonomous systems, inference optimization, RAG, evaluation, and production AI engineering.

Limited slotsGet a Free AI Consultation

How We Work

Custom AI workflows for your Business

One-fit-all AI don't work for modern businesses. At Inferensys, we aim to understand your business & custom requirements; which we use to define most efficient agentic workflows, the data, and the tools for your business.