Mohammad Abdollahi Data Scientist & Operational Researcher

Doctoral research · University of Essex

Using forecast demand to improve delivery decisions

Doctoral research on attended home delivery for online grocery, where forecasts of orders still to come are used to decide which delivery slots to offer, what to charge for them and how to route the vans. Published in the European Journal of Operational Research and Annals of Operations Research.

Illustrative schematic How forecast orders inform delivery-slot decisions A depot and a planned route through confirmed orders. Forecast orders that have not yet been placed are shown as hollow circles along the route. A new customer request can choose slot A, which sits close to the planned route and can be offered cheaply, or slot B, which would require a detour and is therefore priced higher or withheld. Delivery area · one slot window Depot A Slot A: near the route B Slot B: requires a detour Reading the picture Confirmed orders Forecast orders not yet placed Planned route Decision for a new request Slot A adds little cost: offer it at a low charge. Slot B forces a detour and may block forecast orders: charge more, or withhold it. Opportunity cost = expected profit lost on future orders if this request is accepted.
Forecast orders are routed alongside confirmed ones, so the estimated cost of accepting a new request reflects the orders that are still likely to arrive. Positions and outcomes are invented for explanation.

The problem

An online grocer lets each customer choose a delivery slot when they place an order. Every accepted order constrains the routes the vans can drive, and therefore which slots can still be offered to the customers who have not yet ordered. The retailer wants to fill its vans profitably and avoid unnecessary driving, but at the moment it makes each decision it does not know who else will order, where they live or which slot they will want.

The standard tools treat these as separate problems: forecast demand, then price slots, then plan routes. My research connected them. The central idea is to forecast the orders still to come and to include them, provisionally, in the routing plan, so that the estimated cost of accepting any new request reflects the future orders it might displace. That estimate, the opportunity cost of a slot, then drives which slots are shown to a customer and what delivery charge attaches to each.

My contribution

  • Forecast orders inside dynamic routing. Rather than planning routes only around confirmed orders, the routing model carries a set of forecast orders that are progressively replaced by real ones as bookings arrive. This gives a more realistic picture of remaining capacity in each area and slot (European Journal of Operational Research, 2023).
  • Better opportunity-cost estimates. The 2026 paper improves the estimate of revenue lost by accepting a request, using the forecast-based dynamic routing approach, and uses it to set delivery charges.
  • Dynamic slot combination. Many customers are flexible about when they receive a delivery. The 2026 paper introduces an augmented time-window strategy that exploits this flexibility to improve route efficiency without changing the underlying routing system or requiring complex computational modifications.
  • Emissions as well as profit. Because unnecessary driving is a direct environmental cost, the thesis treats reduced mileage as an objective alongside profitability rather than a by-product.
  • A further extension, incorporating the effect of delivery-price reductions into the opportunity-cost estimate, is currently under revision for publication.

How the work was evaluated

The methods were tested in computational experiments driven by real-world customer and operational data from a UK online grocer. Each experiment simulates a booking horizon: customers arrive over time, are shown slots and prices generated by the policy under test, make choices according to a fitted choice model, and the resulting orders are routed. Policies are then compared on profit, delivery efficiency and service.

Across these experiments the proposed policies achieved simulated profit improvements of 13.57–21.43% relative to the comparison policies, and outperformed recent and current state-of-the-art approaches on profitability and delivery efficiency in the 2026 study.

These are simulation results. They show what the policies would have earned under the modelled behaviour of customers and vans on real historical data; they are not a report of profit realised in Ocado’s operations, and the research was conducted independently as doctoral work rather than as a commercial engagement.

What the work produces in practice

For an operator, the outputs are practical rules rather than a single number: which slots to display to a customer at a given address and time, what delivery charge to attach to each, when to combine adjacent slots to gain flexibility, and how to update the plan as orders arrive. A useful property of the approach is that it can be implemented within existing systems that use standard time windows, so a retailer can adopt it without redesigning its routing software.

Limitations

  • Results depend on the fitted customer-choice model; if customers respond to prices differently from the model, the profit estimates shift.
  • Forecast orders are only as good as the demand forecast behind them. The research assumes a reasonably stable booking pattern, which large promotions or disruptions would violate.
  • The experiments cover one retailer’s operating context. Transfer to other delivery settings (parcel, pharmacy, meal kits) is plausible but was not tested.

Why this matters beyond grocery

The same pattern, forecast what is still to come and let that forecast shape the decision you make now, appears in field-service scheduling, appointment booking, ride-hailing and any business that sells constrained capacity over a booking horizon. The research is also the origin of the way I approach applied forecasting work: the forecast is not the product; the decision it supports is.

Discuss a similar problem All work

Also on this site

Industry delivery · Knowledge Transfer Partnership

From demand forecasts to procurement decisions

Designed and built a forecasting and procurement decision system covering 1,200+ products for a UK packaging e-commerce business, as KTP Associate with the University of Kent.