Logistics Intelligence Series | Module 6 of 6

Order-volume forecasting:
What your customers will order next month

A multi-target LSTM learns from 43 customer time series simultaneously and forecasts order volume 4 weeks ahead. The result: proactive capacity planning instead of reactive dispatch.

Model scenario: fictional data

This module is part of a six-part worked example about a fictional mid-sized logistics company (150 trucks, 43 regular customers). All figures, datasets, model metrics and euro amounts are fully simulated. They illustrate the methodology and the order of magnitude of the achievable effect. They are not results of real clients and not a performance promise. What values are achievable on your real data is established only in a pilot project.

01 The problem: reactive dispatch erodes margin

Why "something will turn up" is not capacity planning

A mid-sized logistics company typically plans day by day: orders arrive, trucks are dispatched. Shortfalls are covered from the spot market (expensive). Surplus capacity sits idle in the yard (deadweight cost). Both hurt.

The paradox: the data needed to forecast already exists. Every customer has an ordering pattern: seasonal swings, weekday effects, monthly rhythms. Customer A orders 3 full truckloads every Tuesday. Customer B doubles volume in Q4. Your ERP knows this. Your dispatch team does not.

78% of volume is forecastable

The remaining 22% are genuine spot orders. But forecasting the 78% alone enables internal reallocation instead of spot-market buying at the peak.

Order history
→
Decomposition
→
LSTM-Forecast
→
Capacity plan
→
Dispo-Alert

02 Data foundation: 24 months of order history

43 customers, 52,000 orders, each with volume, timestamp and lane

▸ Example values: method illustration, not a performed analysis
Aufträge: 52.340
Kunden: 43
Zeitraum: 2023-01-02 bis 2024-12-29
Ø Aufträge/Woche: 503
52.340
Orders (2 years)
503
Avg orders / week
43
Regular customers
±22%
Weekly volatility

03 Decomposition: the hidden rhythms

Seasonality, trend and weekday patterns examined separately

Total fleet order volume: 104 weeks + trend
↳ The Q4 peak

Total volume rises 38% in Q4 versus Q2. But not all customers follow this pattern. K-007 (Building materials) peaks in Q2 (construction season); K-031 (E-commerce) surges in November/December. A single fleet model is insufficient. Individual forecasts per customer are required.

Order distribution by weekday (all customers)
Seasonality by industry sector: quarterly comparison

04 LSTM forecast: individual prediction per customer

A recurrent network that learns 43 time series simultaneously

We train a multi-target LSTM: a single model that processes the order history of all 43 customers as parallel time series. The advantage over 43 separate models: it learns cross-industry patterns (e.g. "when Automotive declines, Chemicals rises with a 2-week lag").

▸ Example values: method illustration, not a performed analysis
Modell-Parameter: 148,452
Input:  12 Wochen × 51 Features (43 Kunden + 8 Kontext)
Output: 4 Wochen × 43 Kunden-Prognosen
18%
Fleet MAPE (4-week)
0.891
R² score
22%
Fleet MAPE (week 4)
12%
Fleet MAPE (week 1)
↳ Accuracy by horizon

The fleet-level forecast hits next week within ±12%. Per individual customer the spread remains wider: 12-28% MAPE, driven by the random noise of small order counts.

Forecast vs. actual: total volume (last 16 weeks, test set)
MAPE by customer: 12-28% per customer

The most difficult customers for the model are small e-commerce accounts with high volatility. The best forecasts are achieved for Automotive and Pharma, industries with stable supply chains and predictable rhythms.

05 Customer forecasts: the 4-week outlook

What the weekly capacity report for dispatch would look like

K-031 · E-Commerce · avg. 38 orders/week
↗ Seasonal peak expected
+42% in 4 weeks
Forecast w45-48: 48 → 52 → 54 → 54 orders
Driver: Black Friday run-up + Christmas season. Identical pattern to prior year (±3 orders).
Capacity requirement: +6 truck-days/week additional from w46.
K-007 · Building materials · avg. 28 orders/week
↘ Seasonal decline
-31% in 4 weeks
Forecast w45-48: 24 → 22 → 19 → 19 orders
Driver: End of construction season, falling temperatures. Use freed capacity for e-commerce.
Capacity requirement: -4 truck-days/week from w46. Reallocation to K-031 possible.
K-015 · Pharma · avg. 22 orders/week
→ Stable
±3% in 4 weeks
Forecast w45-48: 22 → 21 → 23 → 22 orders
Driver: Contract-bound volume, low seasonality.
Capacity requirement: No change required. Most reliable planning basis in the fleet.
K-022 · Automotive · avg. 35 orders/week
↗ Trend increase
+18% in 4 weeks
Forecast w45-48: 38 → 40 → 41 → 42 orders
Driver: New model launch at OEM. Not seasonal but a permanent increase for 8 weeks.
Capacity requirement: +3 truck-days/week permanently. Review subcontractor contract.
↳ The dispatch effect

The forecast cards reveal the decisive advantage: K-007 releases capacity that K-031 needs, in exactly the same weeks. Without a forecast, the dispatcher would buy short-term spot capacity for K-031 while K-007's trucks sit idle in the yard. With the forecast: internal reallocation instead of spot-market buying at the peak.

06 Business impact: from forecast to margin

One lever: smoothing spot peaks and idle capacity through better forecasting

€208.800
Total savings / year
0,8%
of the cost base: spot peaks and idle capacity smoothed (assumption)
CategoryAmount/yearMechanismConfidence
Smoothing of spot peaks and idle capacity€208.8000.8% of the €26.1m total cost base (assumption; mechanism anchor McKinsey: AI forecasts reduce errors by 20-50%)Medium

07 The full picture: all 6 modules combined

What the Logistics Intelligence Series means for your fleet

Each module solves a concrete problem. Together they form an operating system for data-driven logistics, built on data you already possess.

Modul 1
€132.300
Dwell Time
Modul 2
€98.820
Tyres
Modul 3
€53.900
Drivers
Modul 4
€243.000
Empty runs
Modul 5
€161.688
Diesel
Modul 6
€208.800
Forecasting
Total potential: €898,508 / year

A model scenario for a 150-truck fleet. The basis: data you already have. Not a single additional customer required.

Savings potential by module: full overview
↳ The message

Around €900,000 lies untapped in data already sitting in your systems. It takes no new ERP and no additional sensors, just the right questions asked of the numbers you already have. That is what we do.

08 Where your forecast goes from here

From proof of concept to live system: the roadmap

You have seen what is possible. The next step is a pilot project with your real data. No months-long concept papers. We deliver results step by step.

① Data workshop (1 day)

We review your telematics, ERP and HR exports together. Which modules are ready to deploy immediately? Where is the highest leverage?

② Pilot

2 modules run on your real data. Deliverable: concrete euro value, validated dashboard, implementation plan.

③ Rollout

Integration into your existing systems. Automated reports, alerts, dashboards. Training for dispatch and fleet management.

All 6 modules: AI in logistics