Your strategic compass for business forecasting. Master the language of predictive analytics with clear, jargon-free definitions designed for decision-makers.
The "Swiss Army knife" of forecasting. ARIMA is a flexible statistical model that combines three components to predict future values: AR (uses past values), I (removes trends to stabilize data), and MA (learns from past forecast errors).
Like a "sugar cane juicer" - the model processes your data multiple times through different components, squeezing out every drop of predictable information until only random noise remains.
Financial data, economic forecasting, sales with trends but minimal seasonality
Moderate
High
A forecasting method that predicts future values based on a weighted combination of past values. The "auto" means it's a regression of the variable against itself.
"The best predictor of what I will do tomorrow is a weighted average of what I did yesterday, the day before, and so on." Captures momentum or persistence in business metrics.
Use Case: Stock prices, GDP growth, any metric where recent performance strongly influences near-term future.
Two ways that trend, seasonality, and noise combine in a time series. Critical for choosing the right decomposition approach.
Seasonal fluctuations stay constant regardless of data level.
Example: +50,000 cases every summer whether baseline is 1M or 5M cases.
Seasonal fluctuations grow proportionally with data level.
Example: +30% holiday sales. At $10M revenue = $3M lift. At $100M = $30M lift.
When to Use: Additive for stable amplitudes, Multiplicative for growing businesses where seasonal peaks expand over time.
Measures the relationship between a time series and a lagged version of itself. It answers: "How much does today's value depend on yesterday's, last week's, or last year's value?" It's the data's "memory."
If today's temperature is high, yesterday's was likely high too (positive autocorrelation = momentum). After buying bulk paper towels, you won't buy more next week (negative autocorrelation = mean reversion).
Why It Matters: Reveals which past periods hold predictive power and helps identify seasonality (e.g., strong autocorrelation at lag 12 means yearly patterns).
Simulating how a forecasting model would have performed in the past by testing it on historical data. Train the model on data up to a specific point, generate predictions, then compare to actual results.
Like putting a pilot through a flight simulator using real historical weather patterns before trusting them to fly a real plane with passengers.
The systematic tendency to consistently over-forecast or under-forecast. It reveals directional errors rather than just error magnitude.
Root Cause: Often caused by organizational behavior (e.g., sales teams "sandbagging" forecasts) rather than model issues.
Medium-to-long-term rises and falls in data that are not of fixed or predictable period. Unlike seasonality, cycles are irregular and often linked to broader economic or business conditions, typically lasting more than a year.
Strategic Value: Distinguishing short-term trends from long-term cycles determines whether to invest aggressively or prepare for downturns.
Challenge of generating forecasts for new products, customers, or markets where no direct historical data exists. Traditional time-series models fail here.
Like a car engine on a freezing day. Existing products (warm engine) run smoothly with their own data. New products (cold engine) need a "jump start" from similar products' data.
Identify "like" products based on attributes (category, price, color) and apply their historical patterns to the new item.
Range likely to contain the true average outcome. Quantifies uncertainty around the mean of predicted values, not any single prediction.
"Based on system-wide data, we're 95% confident our average delivery time next month will be 28-32 minutes." (Strategic question about overall performance)
Use For: Setting company-wide SLAs, broad marketing promises, strategic planning.
Specialized forecasting technique for intermittent demand—sporadic, irregular sales with long periods of zero demand. Common for spare parts, specialized equipment, slow-moving items.
Instead of forecasting demand directly, splits the problem into two simpler components:
Average quantity when sales occur
Average time between sales events
Forecasting ER arrivals for rare injury. Instead of predicting exact minute, predict: "Patients need 3 stitches on average" + "We get one every 72 hours."
Game-changer for: Businesses with long-tail inventory of many slow-moving products.
The collaborative process where leaders from Sales, Marketing, Operations, and Finance review, adjust, and agree on a single, unified forecast that the entire organization uses. Core component of S&OP.
Related: See S&OP (Sales & Operations Planning)
Loss in sales of an existing product caused by promoting or launching a similar product by the same company. Customers switch between your own products instead of generating new demand.
Two Italian restaurants on same street. Massive promotion at "Luigi's Pizzeria" causes sales spike there, but "Mama's Pasta House" sales plummet. Didn't bring new customers—just convinced pasta customers to eat pizza instead!
Overall Lift = Promotional Lift + Halo - Cannibalization - Pull-forward
Breaking down a time series into its four components: Trend, Seasonality, Cycle, and Noise. Acts as a diagnostic dashboard for executives to assess business health.
Seasonal fluctuations stay constant regardless of data level.
Example: +50,000 cases every summer, whether baseline is 1M or 5M cases.
Seasonal fluctuations grow proportionally with data level.
Example: +30% holiday sales. At $10M revenue = $3M lift. At $100M = $30M lift.
A chef tasting a complex sauce and mentally separating core ingredients: underlying sweetness (trend), spice that hits at regular intervals (seasonality), random herb flecks (noise).
A comprehensive business process that goes beyond statistical forecasting. It integrates statistical predictions with qualitative inputs from sales, marketing, and operations to create a unified plan for meeting future demand.
"We predict 70% chance of rain"
"Based on that forecast, we'll bring umbrellas and reschedule the picnic"
Goal: Balance high customer service, low inventory costs, and efficient operations.
Using near real-time data and short-term signals to detect changes in demand patterns quickly. Modern AI can sense demand shifts within hours/days vs. traditional monthly forecasting cycles.
Data Sources: POS data, social media, IoT sensors, web traffic, weather data.
A variable that is determined within the system being modeled. It both influences and is influenced by other variables in the model.
In a sales forecasting model, advertising spend is endogenous because: (1) Ad spend influences sales, BUT (2) Sales performance also influences future ad budgets. It's part of a feedback loop.
Compare with: Exogenous Variable (external factors not influenced by your system)
An external factor that influences your forecast target but is not influenced by it. These are the "outside forces" that affect your business.
Strategic Value: Incorporating exogenous variables transforms forecasting from passive extrapolation to active strategic planning.
A smoothing method that gives greater weight to recent observations. Unlike moving averages where all points in the window are equal, exponential smoothing makes older data exponentially less influential.
For data with no trend or seasonality. Good for stable, mature products.
Adds a parameter to model trend. For data with consistent upward/downward slope.
Handles both trend AND seasonality. Most versatile for business data like holiday retail sales.
Like a rear-view mirror where recent objects appear much larger and more important than distant objects.
The process of predicting future customer demand using historical data, market intelligence, and statistical analysis. Essential for inventory management, production planning, and business strategy.
What Trackura Does: Automates demand forecasting using AI to achieve 85-95% accuracy.
Framework to evaluate whether each step in your forecasting process (statistical model, planner adjustments, manager overrides) makes the forecast more or less accurate.
Forecasting process = assembly line. Raw data enters, each station (model, planner, sales director) performs operation. FVA = quality control checking if each station improved the product or added defects.
Strategic Value: Identify and eliminate steps that destroy value. Simplify process, free up resources, improve accuracy simultaneously.
Creating new, more informative input variables (features) from raw data to enhance a model's predictive power. Using domain expertise to transform raw observations into signals the model can learn from.
Like a chef preparing ingredients before cooking. You don't throw a whole onion in the pot—you peel, chop, and perhaps caramelize it to bring out flavor.
The length of time into the future for which a forecast is generated. Must align with the business decision it supports.
| Horizon | Period | Used For |
|---|---|---|
| Short | 1-4 weeks | Staff scheduling, inventory replenishment |
| Medium | 1-12 months | Sales quotas, marketing budgets |
| Long | 1-5 years | Factory construction, market entry |
Key Factor: Horizon = Lead Time + Planning Cycle frequency
The level of detail at which a forecast is made—from highly aggregated (total category sales) to highly granular (specific SKU at specific store on specific day).
More accurate (noise cancels out)
Best for: Financial planning, strategic decisions
Less accurate but operationally essential
Best for: Inventory ordering, logistics
Methodology for forecasting at multiple aggregation levels (Global → Country → Region → Store → SKU) where forecasts must be coherent—meaning they add up correctly at every level.
Company sales = pyramid. CEO forecasts total at peak. Lower levels = continents, countries, regions, stores. All must add up perfectly!
Strategic Value: Prevents dysfunction where sales team territory forecasts don't match corporate revenue targets. Creates single source of truth.
Demand pattern with sporadic, irregular, infrequent sales—long periods of zero demand interrupted by occasional, unpredictable non-zero events. Also called "lumpy" demand.
Solution: Specialized techniques like Croston's Method that forecast demand size + inter-arrival time separately.
Past values of a variable used as input features for prediction. For example, using yesterday's traffic (lag 1), last week's traffic (lag 7), and last year's traffic (lag 365) to predict today's traffic.
Giving your model a set of rear-view mirrors, each angled to look back at a different specific point in the past. Lag 1 shows what just happened, lag 52 shows this time last year.
Foundation: This is the practical application of autocorrelation analysis—using past periods that have proven predictive power.
Measurable variables that change before the metric you're forecasting, making them powerful early warning signals and predictive inputs.
Like seeing dark storm clouds gathering on the horizon before it starts raining. Gives you time to prepare.
Strategic Value: Provides early warning system for proactive decisions vs. reactive crisis management.
Metrics that change after a business trend has occurred. Used to confirm patterns rather than predict them.
Analogy: Noticing puddles on the ground after a storm has passed (vs. leading indicators = storm clouds before rain).
Examples: Unemployment rate, corporate profits, GDP—they confirm what already happened but don't predict the future.
Applying a logarithmic function to data to compress large values and stabilize variance. Makes exponential growth patterns linear and easier to model.
Business Benefit: Changes in log-transformed data can be interpreted as percentage changes in the original data—highly intuitive for executives.
Use When: Data shows exponential growth (revenue over many years) or is highly skewed (few products = most sales).
Average magnitude of forecast errors in original units (dollars, units, etc.). Answers: "On average, how far off is our forecast?" Less sensitive to occasional large errors than RMSE.
Business Example: MAE of 100 cases means your forecasts are typically off by about 100 cases, whether you predicted too high or too low.
Best For: Operational planning where consistent, predictable error is the main concern.
Average error expressed as a percentage of actual values. Scale-independent, allowing fair comparison across products with different sales volumes.
Best For: Executive reporting and comparing forecast accuracy across different products or business units.
Scale-free error metric that compares your forecast's MAE to a simple Naive forecast's MAE. Provides fair comparison across different time series regardless of scale or seasonality.
Like "Value Over Replacement Player" (VORP). Doesn't measure absolute points scored—measures how many MORE than an average replacement player. MASE measures how much better than a basic Naive forecast.
Ultimate Metric For: Comparing forecast performance across completely different product lines (high-volume screws vs. low-volume machinery). Excellent for intermittent demand.
Forecast created by averaging the most recent data points over a specified "window". Window slides forward for each new forecast. Smooths out noise to reveal underlying patterns.
Like estimating your car's speed by looking only at the last 100 feet of road in your rear-view mirror. Smooths out minor bumps.
Limitation: Gives equal weight to all points in window, so lags behind recent market changes.
Handling data points that are absent from historical records. Critical because ignoring them introduces bias or causes algorithms to fail.
Remove rows with missing data. Only if very few missing.
Fill in missing values with estimates (mean, median, or predicted).
Business Context Matters: Sales missing on Christmas Day isn't random—it's predictable and should be modeled accordingly.
Simplest possible forecast: assumes next period's value will be identical to current period's actual value. Formula: Forecast(t+1) = Actual(t). Also called "random walk" forecast.
"If it ain't broke, don't fix it" approach. Like driving by looking only in rearview mirror—assume road ahead looks like road behind.
Variation: Seasonal Naive = uses value from same period last year instead of last period.
AI models inspired by the human brain, capable of learning extremely complex patterns from vast data. Recurrent Neural Networks (RNNs) and Long Short-Term Memory (LSTM) networks are designed for sequential time series data.
A "black box" super-analyst. Feed it sales, social sentiment, weather, economic data—it learns intricate hidden connections to produce nuanced forecasts.
Use Case: Financial forecasting, complex retail environments, high-stakes problems where marginal accuracy = millions of dollars.
Random, unpredictable variation left after accounting for trend, seasonality, and cycles. Cannot be forecast by definition.
Examples: One-time sales spike from celebrity mention, brief dip from payment processor outage.
Watch For: If "noise" starts showing patterns, it may be a new market factor that needs investigation and modeling.
Putting all input variables onto a common scale (e.g., 0 to 1 range) so variables with larger numbers don't dominate the model unfairly.
Analogy: Converting all recipe ingredients to a single unit (grams) to ensure proportions are correct.
Critical When: Using "number of stores" (small) and "total marketing spend" (millions) in the same model.
Identifying data points that deviate dramatically from the rest. Critical because a single extreme value can severely distort statistical models and pull forecasts in misleading directions.
Like one person shouting in a quiet library. If measuring average noise level, that shout skews the result and gives a false impression of typical conditions.
Open-source forecasting library developed by Facebook (Meta). Designed to be intuitive and easy to use for business analysts, automatically handling seasonality, holidays, and trend changes.
Like a forecasting "smart assistant." Provide your data and important dates (holidays), it automatically generates high-quality forecasts with simple, human-interpretable controls.
The broader field of using data, statistical algorithms, and machine learning to identify the likelihood of future outcomes. Demand forecasting is a key application of predictive analytics.
Scope: Beyond forecasting, includes customer churn prediction, fraud detection, risk assessment, and more.
External factors or variables that help explain and predict your target metric. Provides the "why" behind fluctuations in your data.
Forecasting ice cream sales? Historical sales = your time series. Temperature = your regressor. You look out the window to check weather—it's an external driver of sales.
Types: Exogenous (external: weather, economy) vs. Endogenous (internal feedback loops: ad spend ↔ sales)
Average forecast error in original units, but with heavy penalty for large mistakes. Squares errors before averaging, then takes square root.
Use RMSE when:
Large errors are catastrophic (high-value products, critical inventory)
Use MAE when:
Consistent average accuracy matters more than avoiding occasional big misses
Rule of Thumb: If RMSE >> MAE, you have some very large errors in your forecast that need investigation.
Marketing metric measuring revenue generated per dollar spent on advertising. Forecasting ROAS helps marketers allocate budgets optimally across campaigns and channels.
Trackura Feature: Predict future ROAS before launching campaigns to optimize budget allocation.
Integrated, monthly business management process that aligns all functions (Sales, Marketing, Operations, Finance) to achieve focus and synchronization. Balances supply and demand while aligning operational plans with strategic goals.
Create a single consensus forecast that the entire organization commits to. Prevents dysfunction when different departments operate with different numbers.
Predictable patterns that repeat at fixed, known intervals (daily, weekly, monthly, yearly). Tied to the calendar with consistent timing and magnitude.
Strategic Use: Informs inventory management, staffing decisions, and marketing campaign timing for maximum effectiveness.
Unique identifier for each distinct product variant. Modern forecasting tools can generate SKU-level forecasts for thousands of products simultaneously.
Example: Blue t-shirt, size large, men's = 1 SKU. Red t-shirt, size large, men's = different SKU.
Importance: Supply chain needs granular SKU forecasts, not just "t-shirts"—must know exact sizes/colors/variants to order.
A time series where statistical properties (mean, variance, autocorrelation) remain constant over time. The way data "wiggles" is consistent throughout. Data with clear trends or seasonality is NOT stationary.
Stationary = healthy person's heartbeat at rest (consistent, predictable). Non-stationary = marathon runner's heartbeat (constantly changing, needs context).
Why Critical: Many models (like ARIMA) require stationary data. Analyzing non-stationary data is like navigating a city where traffic laws constantly change.
Method of analyzing data points collected in chronological order at consistent intervals. Fundamentally different from snapshot data—the time dimension reveals how metrics evolve and adjust over time.
What It Reveals: Not just "what happened" but patterns that explain why it happened and when similar events might occur again.
Business Applications: Sales trends, customer behavior tracking, stock prices, employee turnover, website traffic.
Rigorous validation method for time-dependent data that preserves chronological order. Always uses past to train, future to test—preventing "data leakage" where model inadvertently learns from the future.
Train: Year 1 → Test: Year 2. Then Train: Years 1-2 → Test: Year 3. Window grows.
Train: Years 1-2 → Test: Year 3. Then Train: Years 2-3 → Test: Year 4. Fixed size, slides forward.
Long-term, general direction of data—upward, downward, or flat. Reveals fundamental business health independent of seasonal fluctuations or random noise.
Business growth
Decline/concern
Stable/mature
Strategic Value: Distinguish short-term wins from long-term challenges. Strong holiday sales can mask declining underlying trend.
Powerful ML algorithm that builds forecasts by creating an "ensemble" of many simple decision trees. Each tree learns from the errors of previous trees, iteratively improving predictions.
Like assembling a team of specialists. Expert 1 provides general solution. Expert 2 fixes Expert 1's mistakes. Expert 3 fixes remaining mistakes. Final prediction = collective wisdom of entire specialist team.
Use Case: Sophisticated retail sales forecasting where price, promotions, competitors, and economic data all interact.
Put forecasting theory into practice with Trackura's AI-powered platform. Start making data-driven decisions today.