Demand Forecasting in Logistics: Where AI Beats Spreadsheets

Most mid-market logistics operators run demand forecasting in Excel. Templates that mix exponential smoothing with manual adjustments. Year-over-year comparisons updated quarterly. A planning analyst whose primary tool is a spreadsheet she's been refining for three years. The accuracy of this approach is usually somewhere between 60-75% MAPE — meaning average forecast errors of 25-40% per period. ML demand forecasting typically improves this to 80-90% accuracy, which sounds modest but compounds dramatically through inventory, capacity, and labor decisions downstream.

This article is the practical case for moving from Excel to ML demand forecasting in logistics. The technique stack, the integration that determines outcomes, the realistic dollar impact, and where the spreadsheet actually still wins.

What logistics demand forecasting predicts

Three forecast horizons matter for logistics, each used differently:

Short-term (1-7 days). Used for daily dispatch capacity, driver scheduling, and last-mile staffing. The closer to dispatch, the more important real-time signals (weather, current traffic, today's order pace) become.

Medium-term (1-12 weeks). Used for inventory positioning across distribution centers, capacity reservations with carriers, and labor planning. The horizon where ML demand forecasting has the highest leverage.

Long-term (3-12 months). Used for capacity investment decisions, contract negotiations, and network design. The horizon where simpler statistical methods often work as well as ML because the noise dominates the signal.

Most ML investment should target the medium-term horizon — the window where forecasts directly drive operational decisions and where the data history supports the technique stack.

The Excel pattern and why it persists

Excel-based forecasting in logistics has staying power for a real reason: it's transparent, fast to update, and the planning analyst understands every step. When demand surprises happen, the analyst can trace which assumption failed and adjust manually. ML systems often don't have that transparency, which makes operations teams hesitant to trust them.

The Excel approach typically combines:

• Year-over-year baseline (this period last year, with growth adjustment)
• Exponential smoothing on recent demand
• Manual overrides for known events (promotions, holidays, weather forecasts)
• Cross-checks against revenue forecasts from finance
• Final adjustment based on planner judgment

The combination produces forecasts that are explainable, locally adjustable, and trusted. The accuracy is what it is.

Where ML earns its seat over spreadsheets

ML demand forecasting beats Excel-based approaches when:

Volume of SKUs or stops is high. A planner can adjust 50 forecasts manually with judgment. Ten thousand forecasts can't be planner-adjusted; they need a model. ML scales to operations where individual planner attention isn't possible.

Cross-feature interactions matter. Weather × day-of-week × season × promotion calendar creates feature combinations the planner can't manually weight. Tree-based models learn these interactions from data automatically.

External signals are available. Weather APIs, holiday calendars, retail event schedules, macro economic indicators. Models incorporate these systematically; spreadsheets struggle to incorporate them consistently.

Continuous improvement is needed. Models improve as data accumulates. Spreadsheet templates stay roughly the same accuracy quarter after quarter unless the planner manually refines them.

If your operation has all four conditions, ML is a clear win. If you have low SKU count, simple seasonality, and no external signals, the spreadsheet may legitimately be sufficient.

The technique stack that wins

For logistics demand forecasting at most operational scales, the stack is:

Gradient-boosted trees (XGBoost, LightGBM, CatBoost) on engineered features. Lagged demand at multiple horizons, calendar features, weather, promotional calendars, macroeconomic signals. Trains in minutes per model, runs in milliseconds.

Hierarchical reconciliation for multi-level forecasts (SKU → category → total). Models forecast at each level and reconcile so the parts sum to the whole. Useful when both granular and aggregate forecasts matter.

Quantile regression or prediction intervals. Don't just predict the mean; predict the range. A forecast of "100 units, 80% confidence interval [85, 120]" is more useful than just "100 units" because it informs safety stock decisions.

Ensemble across multiple models. Combine ML forecast with statistical baselines (Prophet, ARIMA) for robustness. The combined forecast is usually more accurate than any single approach.

What's NOT in this stack: large language models for the forecasting itself. LLMs are the wrong tool for tabular time-series prediction. They'll produce confident-sounding numbers that are not grounded in your actual data.

The integration that determines outcomes

Forecasting accurately is half the work. Connecting the forecast to operational decisions is the other half — and the half where most logistics ML projects fail.

The forecast must integrate with:

Warehouse Management System (WMS). For inventory positioning across distribution centers. Without this integration, the forecast is academic.

Transportation Management System (TMS). For capacity reservations with carriers and equipment scheduling.

Labor Management System (LMS). For driver and warehouse staffing decisions.

ERP financial planning modules. For budgeting, financial forecasts, and capital decisions.

Each integration takes weeks to months. The integration plan often takes longer than the model build. Plan for it explicitly — model accuracy that doesn't translate into changed operational decisions doesn't generate ROI.

Realistic dollar impact

For a mid-market logistics operator with $100-300M annual revenue moving from Excel-based to ML-based forecasting:

Inventory cost reduction. 5-15% lower carrying cost on the inventory base. For an operation carrying $20M in inventory at 20% carrying cost, that's $200K-600K annual savings.

Capacity utilization improvement. 3-8% improvement in fleet and warehouse utilization. For a $50M fleet operation, that's $1.5-4M in efficiency gains.

Expediting cost reduction. 20-40% reduction in expedited freight and overtime. The most visible savings in the first year because expediting costs show up clearly in budgets.

Service level improvement. 5-10% improvement in fill rate and on-time delivery, which translates to customer retention and pricing power.

The combined impact for a serious deployment is typically 1-2% of revenue annually. The implementation cost (model + integration + training) is typically $200K-800K for a mid-market operator. The math favors investing for any operator above $50M revenue.

What to expect from a real deployment

Realistic timeline:

Months 1-3. Forecast accuracy audit. Identify the baseline. Measure where ML can add value vs where the existing process is sufficient.

Months 4-9. ML model build and integration with the highest-leverage downstream system (usually WMS for inventory or TMS for capacity).

Months 10-15. Deployment to broader portfolio. Integration with secondary downstream systems. Continuous improvement loop.

Year 2+. Mature deployment with monthly model retraining, quarterly forecast accuracy review, and ongoing operational integration.

When to keep the spreadsheet

Three situations where Excel-based forecasting is genuinely the right answer:

Small operations. Below $20M revenue or under 100 SKUs, the planner's judgment plus a spreadsheet often beats ML cost-benefit. The savings don't justify the build.

Operations with very stable demand. A B2B operator with long-cycle contracted demand and minimal seasonality may not benefit from sophisticated forecasting.

Operations where data quality is the limiting factor. If your demand history is unreliable (poor data hygiene, frequent system migrations, inconsistent SKU definitions), invest in data quality before investing in ML. Bad data feeds bad models regardless of model sophistication.

For most mid-market logistics operations, none of these conditions apply, and the move to ML forecasting is a positive ROI investment. For some operations, the spreadsheet stays the right tool — and that's a legitimate finding rather than a missed opportunity.

The honest takeaway

Logistics demand forecasting is the under-marketed AI win. It doesn't show up in the marketing the way route optimization does. The numbers are larger and the integration work is heavier. The teams who treat forecasting as a multi-year operational program build compounding gains across inventory, capacity, and labor. The teams who treat it as a software deployment usually leave the gains on the table.

Match the forecasting investment to the operational decisions it would change. Wire the model into the systems that actually decide. Layer ML on top of the spreadsheet rigor that already works. The pattern that produces dollar gains isn't replacing the planner with AI; it's giving the planner better forecasts and faster cycle times.