Supply Chain & Analytics
Decision Science
There is a family of business problems where the answer is not a prediction but a decision under constraints: which orders to allocate when supply is short, how to schedule a plant, which sites to open, how to assign a workforce. These are optimisation problems, they are solved well by established methods, and they are usually being handled by a spreadsheet and an experienced person's judgement.
That judgement is often good, and it does not scale, does not survive the person leaving, and cannot be tested against alternatives. The value of formalising it is less about beating the human (sometimes the model does not) and more about making the trade-off explicit, testable and repeatable.
How the work runs
Four phases, and what comes out of each.
2–3 weeks
Formulate honestly
The hard part of optimisation is not solving, it is formulating. What is the objective, really, and if there are several, how do they trade off? Which constraints genuinely bind and which are convention that has never been tested? We spend most of this phase separating the two, because a model that treats a soft constraint as hard will produce an answer that is optimal within a fiction.
Outputs
- Objective function agreed, including how competing objectives trade off
- Constraints separated into genuinely binding and merely conventional
- Decision variables and their real operational granularity
3–5 weeks
Build and validate against reality
The model is built in open solvers and first validated by reproducing a period you have already lived through, if it cannot approximately reproduce what actually happened given the same inputs, it does not yet describe your operation. This step catches missing constraints faster than any amount of review.
Outputs
- Working model in open, portable tooling
- Backtest reproducing historical periods within an agreed tolerance
- Missing constraints surfaced by the validation and incorporated
2–3 weeks
Run the scenarios that matter
A single optimum is fragile. We run the decision across demand cases, cost shocks and capacity variations to establish which parts of the answer hold everywhere and which depend on a particular assumption. The robust parts can be committed to; the fragile ones need a hedge rather than a decision.
Outputs
- Scenario matrix with the outcome under each
- Robustness ranking of each element of the decision
- Sensitivity and shadow prices showing which constraint costs most
3–4 weeks
Put it where the decision is made
The model is embedded in the existing decision cadence, presenting a recommendation with its reasoning rather than an unexplained number. Explainability matters more than optimality here, a planner who cannot see why the model recommends something will override it, and they will be right to.
Outputs
- Model integrated into the existing decision workflow
- Recommendations presented with their reasoning and binding constraints
- Override tracking, so systematic disagreement gets investigated
What good looks like
The ranges we design toward.
3–12%
Objective improvement
Typical gain over experienced manual decision-making on allocation and scheduling problems.
Hours → minutes
Decision cycle
Effect of replacing a manual weekly exercise, which also makes scenario testing possible for the first time.
0
Solver licence cost
Built in open solvers. Commercial solvers are faster on very large instances and rarely necessary at this scale.
Deliverables
What you are left holding.
- 01Formulation document separating genuinely binding constraints from convention
- 02Working model in open solvers, handed over and portable
- 03Backtest showing the model reproduces periods you have already lived through
- 04Scenario and sensitivity analysis identifying which decisions are robust
- 05Integration into the decision cadence with override tracking
What an engagement looks like
A representative shape.
Illustrative. This is the shape of work in supply chain & analytics, not an account of a specific client. We do not publish named results, and we would be wary of a firm that publishes them without the client in the room.
12 weeks
- 01Situation
A consumer goods manufacturer was opening a third distribution centre on the strength of a regional sales forecast, with freight cost per unit rising and no cost-to-serve view below channel level.
- 02What we did
Reconstructed eighteen months of shipment data into a true origin–destination model, built cost-to-serve to customer level, and tested the proposed footprint against demand growth, rate shock and node-loss scenarios.
- 03Where it landed
The third site turned out to be robust only under the most optimistic growth case. Redrawing catchments and re-tendering two lanes captured most of the projected saving without the capital, and the site decision was deferred rather than cancelled.
Where this connects
Rarely a self-contained problem.
Questions
The ones we get asked.
How do we know the model is not just wrong in a confident way?
By making it reproduce the past first. If given last quarter's inputs the model cannot approximately reproduce last quarter's decisions and outcomes, it does not describe your operation and no amount of solver sophistication fixes that. This validation step catches missing constraints faster than review, and we do it before anyone is asked to act on a recommendation.
Our planners will override it anyway. Is it worth building?
Only if you treat the overrides as information rather than as disobedience. Systematic override in one direction almost always means the model is missing a constraint the planner knows about, which is a modelling bug surfaced for free. We build override tracking in for exactly this reason. A model whose overrides are never examined will be overridden into irrelevance within two quarters.
Do we need to buy optimisation software?
Usually not. Open solvers handle the great majority of problems at this scale comfortably, and we hand over the model so your team can maintain it. Commercial solvers are genuinely faster on very large instances and buying a licence to answer a one-time question is the wrong order of operations, solve the problem first, then decide whether ongoing use justifies a licence.
What if the optimal answer is operationally unacceptable?
Then a constraint is missing and the model needs it, which is a normal and useful outcome. This is exactly why the formulation phase separates genuinely binding constraints from convention. An answer the operation rejects is information about the model, not a failure of the operation.
Insights
Related reading.
The bullwhip effect: why your demand signal lies
Factory schedules swing by forty per cent while end-consumer demand moves by five. Nobody is behaving irrationally. That is precisely the problem.
Multi-echelon inventory: beyond the safety stock formula
The standard safety stock formula is correct and it is answering a question nobody in a multi-tier network is actually asking.
Cost-to-serve: the analysis most companies skip
Most businesses know their gross margin by customer. Rather fewer know which customers are profitable, and the two lists differ more than anyone expects.
S&OP versus IBP: what actually changes
Most descriptions of the difference are vendor taxonomy. The distinction that matters is whether a trade-off gets settled by someone with the authority to settle it.
Why people systematically under-order high-margin products
The mathematics of a single-period stocking decision is a hundred years old and unambiguous. Experienced buyers still get it wrong in the same direction, every time.
Closely related
More on Supply Chain & Analytics →Forecasting
Models measured against a naive benchmark and backtested honestly, including the finding that some of them should not be deployed.
Supply Chain & AnalyticsAnalytics Operating Model
Where analytics should sit, who owns models in production, and how to stop building things nobody uses.
Supply Chain & AnalyticsNetwork Design
Where to make, where to hold, where to ship from, settled with a cost-to-serve model instead of an argument.
Start a conversation
Talk to us about decision science.
Most useful first conversations are thirty minutes and involve no proposal. Describe the problem you are looking at and we will tell you honestly whether it is one we can help with.
Or write directly to hello@shyenova.com
