When it comes to supply chain capacity planning, the forecast is almost always wrong. That's not new. What is new is how spectacularly off it's been over the last few years and how many companies are still clinging to approaches designed for a world that no longer exists.
The pandemic. The Suez Canal blockage. Semiconductor shortages cascading across dozens of industries. Demand whiplash from panic buying to inventory gluts in under eighteen months. Traditional demand forecasting wasn't built for this. And pretending otherwise is costing companies billions.
So let's talk about what actually breaks down, and more importantly, what actually holds up.
The problem with "normal"
Classical demand forecasting relies on a fundamental assumption: the future will resemble the past. Take historical sales data, apply statistical smoothing, layer in known events like promotions or holidays, and project forward. This works beautifully in steady-state environments with stable demand patterns, predictable lead times and reliable supplier networks.
But disruptions don't care about your baseline. When a shock hits, whether supply-side constraint, sudden demand spike, or structural shift in consumer behavior historical patterns become noise, not signal. The lag between "the world changed" and "our forecast caught up" is where the real damage happens: stockouts, excess inventory, misallocated capacity, and frantic firefighting that burns out planning teams.
A forecast that looked reasonable on Monday becomes laughably wrong by Thursday because a key supplier declared force majeure. Or, demand for a category tripled in a week because of a viral social media moment that no time series model could have predicted.
The uncomfortable truth? In volatile environments, a sophisticated statistical forecast can perform worse than a simple naive model because it's confidently wrong. It gives you false precision when what you need is honest uncertainty.
What actually holds up
Scenario planning (done right)
The companies that navigated recent disruptions best weren't the ones with the most accurate point forecasts. They were the ones who had already asked: "What if demand is 40% above plan? What if a key supplier goes down for six weeks?"
Good scenario planning means defining three to five plausible scenarios, quantifying operational implications of each, pre-deciding trigger points and response playbooks, and updating scenarios regularly. The output isn't a better forecast - it's preparedness.
Shorter planning cycles
Most companies plan on monthly or quarterly cycles because that's how their S&OP process works. But in volatile environments, monthly plans are stale on arrival. Leading organizations are moving toward weekly planning cycles for their highest-variability segments by layering a faster sensing and response loop on top of the annual plan.
This doesn't mean throwing out your annual plan. It means giving planners the authority and the tools to adjust within guardrails rather than waiting for the next planning cycle to officially acknowledge what everyone already knows.
The biggest barrier isn't technology. It's culture - the discomfort of admitting that last week's plan is already outdated.
ML augmentation (with realistic expectations)
Machine learning can genuinely improve demand forecasting, but not in the way vendors often sell it. Where ML helps: incorporating diverse signals (web search trends, weather, social media velocity, competitor pricing), pattern recognition at scale across thousands of SKUs, and automated anomaly detection when actual demand deviates from forecast.
Where ML doesn't help: true black swan events with no historical precedent, situations where training data is from a fundamentally different regime, and when the real constraint is supply rather than demand. The right framing: ML as a copilot for your planning team, not a replacement.
Demand sensing
Demand sensing uses near real-time data point-of-sale data, order pipeline trends, channel inventory positions, web traffic to adjust short-term forecasts based on what's actually happening now rather than what was predicted months ago. Its power isn't predicting disruptions; it's detecting them faster, shortening the gap between "something changed" and "our plan reflects that change."
The caveat: demand sensing works best for the near-term horizon (1-4 weeks out). It doesn't replace the need for longer-range forecasting for capacity decisions and supplier commitments. It's a layer, not a replacement.
Practical takeaways
• Measure forecast value-add, not just accuracy. If your forecast isn't better than a naive model during volatile periods, invest in responsiveness instead.
• Separate what you can forecast from what you can't. Plan with ranges and scenarios for unpredictable segments, not point estimates.
• Invest in sensing speed over prediction accuracy. Detecting a demand change within days is often worth more than a marginally better forecast made weeks in advance.
• Build optionality into your supply network. The best forecast improvement is often a supply chain that can flex without needing a perfect forecast.
• Get comfortable communicating uncertainty. Give leadership ranges, probabilities, and scenarios rather than a single number you don't believe in.
The era of low-volatility, predictable global supply chains is likely behind us. The companies that thrive won't be the ones with the best models; they'll be the ones with the fastest learning loops, the most honest relationship with uncertainty, and the organizational agility to act on imperfect information. That's what actually works.
