Demand forecasting and inventory management consulting

How you forecast today

We start by understanding your process: how forecasts are produced, what they are used for, and by whom. Which decisions depend on them, where judgement enters, and where the numbers get overridden. This tells us what already works and what is worth changing.

Whether it holds up

Then we measure it properly. How accurate your forecasts really are, and whether your error measures reflect the decisions they feed. Where demand is censored by stockouts, so your history understates what customers actually wanted. How your safety stocks are set, and whether they match the service levels you are targeting.

What it takes

A month of analysis and writing. From you: sales and stock data with the longest history you have either in Excel, CSV, or as direct access to your database.

Cost depends on the scale of the problem — a thousand SKUs is a different exercise from a hundred thousand, and a clean database export is different from a folder of spreadsheets. We scope it once we know what situation you have.

Stock you do not need

Safety stock set from the wrong numbers (or from a rule of thumb) ties up working capital in items that were never going to sell that fast. Correcting the calculation typically releases cash without touching service levels.

Sales you cannot make

The other failure mode, and the more expensive one: the item a customer wanted was not there. Lost sales typically do not appear in reports, which is why they persist.

The service level nobody chose

Most companies target a service level by convention rather than by calculation. In a recent analysis, moving from a 99% target to 90% roughly doubled profit on the items concerned, and a 99% target service level was loss-making across an entire product group — the inventory bill exceeded the margin it earned.

Planner time

When the system is not trusted, planners override it by hand. Most of those overrides make the forecast worse, and all of them cost time that a properly specified model would give back.

Where the forecast was being undone by hand

A branded consumer goods manufacturer, around 1,500 products. We audited the forecasting process end to end and benchmarked the company’s own forecasts against nine suitable statistical methods.

The forecasts were adequate, but more than half of them were being adjusted manually. Separating the adjustments by size showed a clear pattern: moderate adjustments improved accuracy by around 58%, while adjustments of more than 2,000% made the forecast roughly twenty times worse. Most of the beneficial adjustments were planners correcting for promotions — something the statistical model could have handled directly, had it been given the promotional data.

Delivered: process audit, ABC-XYZ categorisation, and a set of rules for where judgement should and should not be used.

Testing a software vendor’s own engine

An analytics software company asked us to benchmark the forecasting approach inside their product against established methods, on nearly 20,000 intermittent retail series.

Their method was beaten on point accuracy by Croston and TSB, well established statistical methods that run in a fraction of the time. More seriously, their prediction intervals were badly miscalibrated: intervals sold as 85% covered only 73% of actual values, and some quantiles came back negative, which is unsuitable for demand. Such narrow intervals look good until they are translated into safety stock, at which point they lead to stockouts.

Delivered: benchmarking study, diagnosis of the interval problem, and a forecasting engine improvement roadmap.

What the stock policy was actually worth

An industrial distributor with a catalogue of roughly 33,000 SKUs, running replenishment through an ERP system. Demand was almost entirely intermittent: 82% of product-weeks had no sales at all.

We tested ten forecasting approaches (including statistical and ML), simulated an order-up-to policy at four service levels, and calculated costs of every strategy in the client’s own currency. The model that minimised costs turned out to be the wrong answer — it saved money by heavily understocking. Once profit replaced cost as the objective, a simple 12-week moving average at a 90% target service level won almost across the board, while a 99% target proved loss-making.

Delivered: a per-product recommendation file covering every active SKU, and a script that regenerates safety-stock targets whenever new data arrives.

Training your team

Often the highest-value next step. Many forecasting problems are not tooling problems but understanding problems: the methods are available, but nobody was taught when to use which. We run practitioner courses and bespoke training for planning teams.

Implementation

Where the recommendations need building rather than explaining. We develop forecasting and inventory models around your demand patterns, your drivers, your ERP, and your planning cycle — and make sure the output is something your planners can actually run each week.

The right method, not the fashionable one

We use machine learning where it earns its place, and statistical models where they do. Neural networks, gradient boosting and conformal methods beat conventional approaches on some problems and lose badly on others, particularly with short histories or sparse demand. Knowing which case you are in is the expertise.

Methods you can inspect

Our work runs on the open-source packages we build and maintain — smooth and greybox — grounded in a published methodology (ADAM) and benchmarked on international forecasting competition data. You can read exactly what our methods do before you hire us.

Explainable models

You can see what drives every number: which components the model uses, how much each demand driver contributes, and why the forecast changed since last month. That matters when a planner overrides a figure, when finance asks why the plan moved, or when a forecast has to be defended in a meeting. Models you cannot interrogate get quietly ignored, and then the process runs on judgement alone.

No software to buy

We are not reselling a platform. The tools are free and open source, so what you pay for is the analysis and the judgement, not a licence. Nothing we build locks you into us.