Time Series Forecasting Basics: A Guide for Demand Planners

If you have ever stared at a sales chart trying to guess where the line goes next, you have performed time series analysis.

In the world of supply chain, time series data is the heartbeat of your operation. It is the sequence of sales, inventory levels, or shipments collected at regular intervals—daily, weekly, or monthly.

But raw data can be deceptive. A spike in sales might be a trend, a seasonal holiday rush, or just random noise. Misinterpreting these signals leads to the two nightmares of demand planning: excess inventory or stockouts.

You don't need a PhD in statistics to build a reliable forecast, but you do need to understand the anatomy of your data. In this guide, we will break down the components of time series forecasting, compare the most common methods, and help you choose the right approach for your business.

What is Time Series Forecasting?

Time series forecasting is the technique of using historical data to predict future values. Unlike causal models, which might look at "why" something happened (e.g., "sales went up because we lowered the price"), pure time series forecasting looks at "what" happened in the past to project the future pattern.

It relies on the assumption that history, while it doesn't repeat exactly, often rhymes.

The Anatomy of a Time Series

To fix a forecast, you first need to diagnose the data. Think of a time series like a signal composed of four distinct layers. This is often called "decomposition."

1. Trend

The trend is the long-term direction of the data. Is your product growing, declining, or flatlining?

• Example: A steady 5% year-over-year growth in your core SKU.
• Practitioner Note: Don't confuse a short-term spike with a long-term trend. Trends usually play out over many periods.

2. Seasonality

These are patterns that repeat over a fixed period—a week, a month, a quarter, or a year.

• Example: Ice cream sales peaking in July, or B2B software sales spiking at the end of Q4.
• Practitioner Note: Seasonality is predictable. If you know it happens every December, your model should expect it.

3. Cyclicality

Cycles are longer-term waves that don't have a fixed period. These are often tied to the wider economy.

• Example: A recession reducing demand for luxury goods for 18 months.
• Practitioner Note: Unlike seasonality, cycles are hard to predict purely from internal data.

4. Noise (Residuals)

This is the random variation left over after you account for trend, seasonality, and cycles.

• Example: A blizzard delaying shipments, or a competitor running a surprise flash sale.
• Practitioner Note: You cannot forecast noise. If you try to fit your model to the noise (a common mistake called overfitting), your forecast accuracy will plummet.

Visualizing the Story

Before you run a single algorithm, plot your data.

A simple line chart often reveals the obvious. Are there massive gaps? (See "Ghost Demand" below). Is there a clear seasonal heartbeat?

A common technique is Seasonal Decomposition, which splits your chart into three separate graphs:

1. Observed: The raw data.
2. Trend: The smooth underlying movement.
3. Seasonal: The repeating pattern.
4. Residual: The noise.

If the "Residual" chart looks like a random scatter of dots, your model is good. If there are patterns in the residuals, your model missed something.

The "Stationarity" Check

You might hear data scientists talk about "stationarity." In plain English, a stationary time series is one whose statistical properties (mean and variance) don't change over time. It's a flat line with some wiggles.

Most sales data is non-stationary—it trends up and down and has seasonality.

Why does this matter?

• Some models (like ARIMA) require you to make the data stationary first (usually by "differencing"—subtracting today's value from yesterday's).
• Other models (like Holt-Winters) handle non-stationarity naturally.

If you are using a modern tool like DemandPlan, this transformation often happens in the background. But knowing why a model fails on trending data is a key skill.

Common Forecasting Methods Explained

You have dozens of algorithms at your disposal. Here are the workhorses of the industry.

1. Naive Method

The simplest approach: "Tomorrow will be exactly the same as today."

• Use Case: A benchmark. If your complex AI model can't beat the Naive forecast, you have a problem.

2. Moving Average (MA)

Takes the average of the last n periods.

• Use Case: Stable products with no trend or seasonality. It smooths out noise but lags behind trends.

3. Exponential Smoothing (ETS / Holt-Winters)

This is the gold standard for many demand planners. It assigns weights to history, giving more importance to recent data. Holt-Winters is a version that specifically handles Trend and Seasonality.

• Use Case: Seasonal products (e.g., retail, seasonal food and bev). It is robust, explainable, and fast.
• Learn more: Deep dive into Exponential Smoothing.

4. ARIMA (AutoRegressive Integrated Moving Average)

A more complex statistical approach that looks at autocorrelations—how today's value correlates with values from 1, 2, or 12 periods ago.

• Use Case: Data with complex patterns that aren't purely seasonal. It requires more data history than ETS to be stable.

5. Machine Learning (ML)

Models like XGBoost or LSTM that can ingest external features (price, weather, promotions) alongside the time series history.

• Use Case: When history alone isn't enough. If you run a promotion, history says "sales are normal," but ML knows "price drop = sales spike."

How to Choose the Right Method

There is no "one ring to rule them all." The best method depends on your data's profile.

| If your data has... | Consider this method | | :--- | :--- | | Short history / New Product | Moving Average or "Like-Item" modeling | | Strong Seasonality | Holt-Winters (ETS) | | Complex Trends + Seasonality | ARIMA or Prophet | | External Drivers (Price, Ads) | Machine Learning / Regression |

For a comprehensive breakdown, check our guide on Forecasting Methods.

Common Pitfalls

Overfitting

This is when you force a model to memorize the past so perfectly that it fails to predict the future. It's like studying for a test by memorizing the exact answers to the practice exam, rather than learning the concepts.

• The Fix: Always test your model on "held-out" data (data the model hasn't seen).

Ignoring "Ghost Demand"

This is a classic supply chain trap. Your sales history records what you sold, not what customers wanted. If you were out of stock for a week, your sales were zero, but demand was likely high.

• The Fix: You must correct your history for stockouts before feeding it into a time series model. Otherwise, the model will learn that "demand drops to zero every few months" and forecast a drop exactly when you need inventory most.

How DemandPlan Approaches It

At DemandPlan, we believe in a "Hybrid" approach.

1. Statistical Baseline: We use robust methods like ETS to generate a solid baseline forecast based on history.
2. AdaptiveHierarchy™: We aggregate time series data. Sometimes a SKU is too noisy to forecast alone, but the Category has a clear trend. By looking at the hierarchy, we stabilize the signal.
3. Human Adjustment: We make it easy for you to overlay market intelligence (like a new customer contract) on top of the statistical baseline.

Conclusion

Time series forecasting is part art, part science. By understanding the components of your data—Trend, Seasonality, and Noise—you can stop guessing and start planning.

Start simple. Visualise your data, establish a baseline with a method like Holt-Winters, and then layer on complexity only where it adds value.

Ready to improve your forecast accuracy? Understanding the math is step one. Step two is measuring your results. Read our guide on Forecast Accuracy Metrics to see how to keep score.

Or, if you are tired of fighting with spreadsheets, see how DemandPlan automates this process.