Understanding Demand Planning Without the Hype
Demand planning is the process of predicting what customers will buy and when they'll buy it. It sounds straightforward. It isn't. The people who treat it like a math problem end up with spreadsheets that look good but mean nothing on the floor. I've watched it happen more times than I care to count. The fundamentals break down into a few core areas: understanding your data, choosing the right forecasting method, building a baseline model, and then layering in adjustments for known events. Jack's approach seems to focus on making these steps feel less like academic exercises and more like something you can actually run in a company that's already behind on its quarterly targets. Most beginners skip straight to the software. They install whatever demand planning tool the VP of Supply Chain bought and feed it ten years of historical sales. That won't work. You need clean data first. You need to know which SKUs are actually sellable, which rows in your system are returns that got double-counted, and which regions have seasonal patterns that the rest of the company treats as noise.
The Methods That Actually Work in Practice
There are three buckets you need to know about. Quantitative methods use historical data. Qualitative methods rely on expert judgment, surveys, or market research. And hybrid approaches combine both. In the real world, you're always using at least two of these simultaneously, even if your software interface pretends otherwise. Moving average models are the entry point. They smooth out short-term volatility by averaging sales over a rolling window. Simple enough. Exponential smoothing adds weight to recent observations, which matters when your product lifecycle is shorter than twelve months. If you're forecasting consumer electronics or fashion, naive averaging will make you miss demand shifts by three to four weeks. Causal models like regression analysis link your demand to external variables. Price changes, promotional spend, competitor activity, weather patterns — anything that actually moves the needle. These are harder to build but dramatically more accurate when the relationships are stable. The problem is that relationships aren't stable. I worked on a project last year where we'd spent six weeks building a regression model for a beverage client, only to find that their primary distributor had quietly switched logistics providers mid-quarter. All the price elasticity coefficients became garbage overnight. We ended up falling back to a simple exponential smoothing model with a manual override for the distribution event. It was uglier but it didn't lie to them.
Where People Mess Up
The biggest mistake I see is treating the forecast as a single number. It should always be a range. A point estimate like "we'll sell 14,327 units next month" is a false sense of precision. What you need is a prediction interval: "we'll sell between 12,100 and 16,800 units, with roughly a sixty-eight percent confidence level." That tells your operations team whether to run one shift or two. That tells procurement whether to negotiate a firm contract or keep options open. Another common error is ignoring the sales and operations planning process. Demand planning doesn't exist in a vacuum. If your sales team is running promotions you don't know about, if your marketing is launching products you haven't been told about, your forecast will drift. I've seen forecast accuracy drop from around eighty-five percent to below sixty percent in a single quarter because the demand planning team wasn't in the room when the product launch was approved. The fix is simple in theory. Embed demand planners in the S&OP meeting rhythm. Make sure every new initiative gets flagged before it hits the market.
Get the Full Details
Building Your First Model Step by Step
Start with your data. Gather at least twenty-four months of historical demand at the SKU-level, preferably broken down by region or channel. Remove outliers caused by stockouts, returns, or one-time bulk orders. A stockout isn't demand — it's a missed opportunity, and if you include it as zero sales in your history, your model will learn to underforecast. Next, choose your base method. For stable, mature products with consistent demand patterns, start with double exponential smoothing (Holt's method). For products with clear seasonality, use Holt-Winters. For anything with known drivers like promotions or price changes, run a multiple regression alongside your time series model and compare the results. Then validate. Split your data into a training set and a holdout set. A common approach is to use the first eighty percent for training and the last twenty percent for testing. Calculate your MAPE (Mean Absolute Percentage Error) and your WAPE (Weighted Absolute Percentage Error). WAPE is generally better because it doesn't blow up when actual demand is near zero. If your model is giving you errors above thirty percent, something is wrong. Maybe the wrong method. Maybe dirty data. Maybe a structural change in the market you haven't accounted for.
The Hard Truths About Accuracy
Forecast accuracy has diminishing returns. Getting from zero to seventy percent accuracy is relatively easy. Getting from seventy to eighty is harder. Getting from eighty to ninety is expensive. Most companies stop optimizing once they hit the point where the cost of further accuracy improvements outweighs the benefit. That threshold varies wildly by industry. For grocery retail, a one-percent improvement in forecast accuracy might save millions in reduced waste and stockouts. For industrial equipment manufacturers selling custom machinery on six-month lead times, the margin is thinner and the forecast horizon is longer, which changes the entire calculus. Also, accuracy means different things at different levels of aggregation. Your forecast might look great at the category level but terrible at the SKU-level. Aggregation reduces variance. That's a feature, not a bug. Don't get hung up on SKU-level perfection. Focus on getting the aggregate right, then deal with individual line items through manual review or safety stock adjustments.
When the Basics Break Down
Sometimes historical data simply doesn't exist. New product launches, discontinued products, market entries into new regions. In those cases, you're not forecasting. You're estimating. Use analogies from similar products. Run market research. Talk to your sales team. Accept that the error margin will be wide and plan your inventory accordingly. Carrying extra safety stock for unknown demand is cheaper than stocking out on a new product that could have been a hit. Jack's approach to these edge cases tends to emphasize transparency — showing the uncertainty explicitly rather than pretending the model knows more than it does. That's sound advice. The worst forecasts are the ones that look precise but are built on assumptions nobody questioned.
