DO Radar — 29 Apr 2026

🎧 The Brief

Industry Signals

Industry Signal Energy FICO Xpress MILP Apr 2026

AFRY BID3 Wins 2026 FICO ESG Champion Award: FICO Xpress MILP Powers Energy Market Bidding and Unit Commitment for 30 Global Utilities, 25% User Growth in 2025 🔗

Foundation: Electricity markets operate on a two-settlement structure — day-ahead and real-time — in which generators must commit hours or days before delivery, locking in which units will run, at what output level, and at what bid price. The decision is a Unit Commitment (UC)The power-system scheduling problem that decides which generators to turn on or off each hour to meet demand at minimum cost, subject to ramp rates, minimum up/down times, and reserve constraints. problem: a large Mixed Integer Programme (MIP)A linear programme with some variables restricted to integer values; the workhorse of discrete optimisation. that must honour physical constraints (minimum up and down times, ramp rates, spinning reserve requirements, capacity limits) while minimising system cost or maximising generator profit. For a single national grid with hundreds of generators and 24 hourly time periods, the resulting MIP has tens of thousands of binary variables and can take hours to solve with a general-purpose branch-and-bound solver. Energy consultancies and market participants that advise governments on grid investment and market bidding strategy need to solve this MIP repeatedly under different scenario assumptions — regulatory change, renewable build-out, demand growth, carbon-price trajectories — to answer the questions that underpin billion-dollar infrastructure commitments.

AFRY, the global engineering and infrastructure consultancy, built BID3 — a power market modelling suite — specifically to solve this problem at national and regional scale. BID3 applies Mixed Integer Programme (MIP)A linear programme with some variables restricted to integer values; the workhorse of discrete optimisation. via the FICO® Xpress Optimization solver to day-ahead and intra-day energy market bidding, generator unit commitment scheduling, and long-term capacity planning across interconnected grids. FICO Xpress is explicitly named as the computational engine; AFRY describes it as the layer that transforms a policy question into a tractable optimisation problem and back into an actionable investment recommendation. BID3 now serves 30 organisations worldwide — grid operators, utilities, policymakers, and investors — and recorded 25 % year-on-year user growth in 2025 as demand for sophisticated energy-transition modelling accelerated. AFRY positions BID3 as a “GPS for net-zero”: the system identifies cost-effective decarbonisation pathways while maintaining energy security and affordability constraints, and it has informed national energy policy decisions in multiple countries. On 18 April 2026, AFRY’s BID3 platform was named the winner of the 2026 FICO Decision Award for ESG Champion category.

Why it matters: The AFRY BID3 deployment illustrates two things practitioners should take from this week. First, Cutting PlanesValid linear inequalities added to an LP relaxation to tighten it toward the integer hull, reducing the gap between the LP bound and the optimal MIP solution. implemented inside FICO Xpress are what makes the unit commitment problem solvable at planning speed: without Branch-and-CutAn exact MIP algorithm that builds a search tree of LP relaxations, pruning branches proven suboptimal and adding cutting planes to tighten the LP bound., the LP relaxation at each branch-and-bound node would be far too loose to prune the search tree efficiently, and a 30-organisation multi-scenario model would be intractable. Second, the 25 % user growth figure signals that energy-sector demand for provably-optimal MIP tooling is expanding, not contracting, even as reinforcement-learning and forecast-only approaches attract more headlines. For practitioners in energy markets, the deployable pattern is to separate the forecast layer (renewable generation, demand, fuel prices) from the optimisation layer (unit commitment and bidding) and connect them with a commercial MIP solver; BID3 shows that this separation enables billion-dollar decisions to rest on reproducible, auditable mathematics rather than heuristics.

Source: FICO Decision Awards 2026 · ESG Champion · 18 Apr 2026

Industry Signal Financial Services MILP Pricing 📋 Case Study Apr 2026

Erste Group Deploys FICO Decision Optimizer for Retail Lending Across Central and Eastern Europe: 22% Profit Lift, 90% Manual Exception Reduction — 2026 FICO AI/ML/Optimisation Award 🔗

Foundation: Retail loan pricing in a commercial bank is an optimisation problem that almost no bank solves as an optimisation problem. The naive approach sets a rate schedule by product type and adjusts at the branch level, giving relationship managers discretion to match competitor offers or reward loyal customers. In practice, this produces a pricing distribution that is driven by which manager a customer happens to visit: nearly 90 % of small business loan prices at Erste Group were set by branch-level exceptions rather than any systematic rule. A Mixed Integer Programme (MIP)A linear programme with some variables restricted to integer values; the workhorse of discrete optimisation. pricing model instead sets loan prices that maximise portfolio-level profitability subject to regulatory capital constraints, credit-risk appetite, competitive positioning bounds, and customer-level fairness constraints. The solver decides, for each loan application, the interest rate that maximises expected contribution to the portfolio objective given the customer’s risk profile and the bank’s current balance-sheet position — not the rate a branch manager negotiated on a given Tuesday afternoon.

Erste Group, one of Central and Eastern Europe’s largest retail banks, has deployed FICO Decision Optimizer — a Mixed Integer Programme (MIP)A linear programme with some variables restricted to integer values; the workhorse of discrete optimisation.-based pricing and portfolio optimisation platform — for loan pricing across mortgages, cash loans, and small business unsecured installment loans in multiple markets. The initiative replaced the exception-driven system with systematic, solver-determined pricing and delivered a 22 % increase in loan portfolio profitability in the small business installment loan segment. Before the deployment, the bank’s pricing was effectively decentralised: managers set individual rates, the portfolio had no aggregate objective, and profitability was unmeasurable at the transaction level. FICO Decision Optimizer replaces that with a single optimisation that prices every loan against an explicit portfolio objective, with Cutting PlanesValid linear inequalities added to an LP relaxation to tighten it toward the integer hull, reducing the gap between the LP bound and the optimal MIP solution. used internally to tighten the solver’s LP relaxation and keep solve time compatible with real-time decisioning. Erste Group has developed this capability over nearly 14 years and was named the winner of the 2026 FICO Decision Award for AI, Machine Learning and Optimisation on 18 April 2026.

Why it matters: Erste Group’s deployment is the clearest published evidence that retail banking loan pricing is a tractable Mixed Integer Programme (MIP)A linear programme with some variables restricted to integer values; the workhorse of discrete optimisation. problem with a 22 % profitability improvement at stake. For practitioners in financial services, the deployable lesson is the exception elimination: moving from a 90 % exception rate to a systematically optimised price does not require eliminating banker discretion — it requires framing discretion as a constrained variable within the solver rather than an unconstrained override of it. The 14-year deployment timeline also illustrates a frequently overlooked dynamic in financial MIP applications: regulatory approval, model validation, and credit-risk governance add 2–5 years to the deployment cycle even when the mathematical solution is ready, so organisations that start the regulatory engagement early have a structural advantage over those that treat it as a downstream compliance step.

Source: FICO Newsroom · Erste Group · Apr 2026📋 Case Study below ↓

► 📋 Case Study production deployment — FICO Newsroom + 2026 Decision Award (Apr 2026) Expand for full case ▼

Takeaway

Erste Group replaced 90 % manual exception pricing with a MILP solver and lifted loan portfolio profitability by 22 % in the small business segment. ↓ Expand to see formulation, difficulty, and full takeaway

Problem

Erste Group operates retail lending across multiple Central and Eastern European markets. Loan prices (interest rates) were set by a combination of rate schedules and branch-manager exceptions. Nearly 90 % of small business unsecured installment loan prices were driven by branch-level exceptions — individual negotiations rather than any portfolio-level objective. The result was a pricing distribution that reflected relationship dynamics rather than risk-adjusted profitability, making it impossible to measure or improve portfolio-level contribution at the transaction level.

Formulation

Type Mixed Integer Linear Programme (MILP) for loan pricing optimisation

Objective Maximise portfolio-level expected profitability across all active loan applications

Decisions Interest rate assigned to each loan application (discretised to pricing tiers)

Constraints Regulatory capital requirements (Basel III/CRR); credit risk appetite (maximum expected loss); competitive pricing bounds (rate floors and ceilings by product); fairness constraints (rate dispersion limits across customer segments)

What made it hard

Three compounding difficulties. First, the pricing model must be approved by bank regulators before it can influence customer-facing decisions, adding 2–5 years of model validation and governance overhead to the deployment timeline. Second, the objective function (portfolio profitability) is a non-linear function of individual loan prices when credit risk correlations are included; linearising it for MILP requires careful approximation that must be validated against full non-linear benchmarks. Third, the 14-year development timeline reflects the need to build internal organisational capability alongside the model — pricing decisions made by a solver rather than a banker require extensive stakeholder alignment across credit risk, compliance, and distribution teams.

What was done

Erste Group deployed FICO Decision Optimizer, a MILP-based pricing platform, to set loan prices systematically across its retail lending portfolio. The platform takes each loan application as input, evaluates it against the portfolio-level pricing objective and all regulatory constraints, and outputs a recommended price. For small business unsecured installment loans, the deployment replaced the exception-driven pricing process entirely: the solver sets the price and branch managers work within a tighter, solver-defined range. The solver uses cutting planes internally (as part of the FICO Xpress branch-and-cut engine) to handle the integrality constraints on pricing tiers and the binary variables representing product eligibility. The result was a 22 % increase in profitability in that segment and a significant reduction in manual exceptions.

Practitioner Takeaway

The Erste Group case makes two things concrete for financial services practitioners. First, the 22 % profit improvement came from eliminating exception-driven pricing, not from a novel algorithm — the MILP formulation is standard, the value came from applying it consistently at scale. Second, the 14-year timeline is not a failure; it reflects the genuine cost of regulatory approval and organisational change in a regulated industry. Practitioners who want to apply MILP pricing in banking should start the regulatory engagement at the design phase, not at the deployment phase.

Research Papers

Research Paper Energy MILP + Deep Learning arXiv · 23 Apr 2026

A Multi-Stage Warm-Start Deep Learning Framework for Unit Commitment 🔗

Foundation: Unit Commitment (UC)The power-system scheduling problem that decides which generators to turn on or off each hour to meet demand at minimum cost, subject to ramp rates, minimum up/down times, and reserve constraints. is a large-scale Mixed Integer Programme (MIP)A linear programme with some variables restricted to integer values; the workhorse of discrete optimisation. with tens of thousands of binary variables representing on/off decisions for each generator in each time period. Commercial solvers like FICO Xpress use Cutting PlanesValid linear inequalities added to an LP relaxation to tighten it toward the integer hull, reducing the gap between the LP bound and the optimal MIP solution. to tighten the LP relaxation and branch-and-bound to search the integer space, but even with these tools a full solve can take hours on large grids. Grid operators need solutions in minutes — not hours — to respond to real-time changes in demand and renewable generation. The standard workaround is a time limit with optimality gap tolerance, which works but leaves gap-closing potential on the table. A different approach: use a fast ML model to predict a good initial solution (warm start), then let the MILP solver close the remaining gap from a much better starting point.

Researchers posting to arXiv on 23 April 2026 (arXiv:2604.21891) propose a multi-stage pipeline for Unit Commitment (UC)The power-system scheduling problem that decides which generators to turn on or off each hour to meet demand at minimum cost, subject to ramp rates, minimum up/down times, and reserve constraints. that uses a transformer-based deep learning model to predict generator commitment schedules over a 72-hour horizon and feeds those predictions as a warm start to a downstream Mixed Integer Programme (MIP)A linear programme with some variables restricted to integer values; the workhorse of discrete optimisation. solver. The pipeline applies post-processing heuristics to enforce physical constraints (minimum up/down times, ramp rates) before the MILP solver runs, and uses a confidence-based variable fixation strategy to fix high-confidence binary decisions and reduce the combinatorial search space. The full multi-stage pipeline achieves 100 % feasibility — every instance produces a valid commitment schedule — and in approximately 20 % of test instances the ML-warmed solver reaches a lower overall system cost than the solver alone within the same time budget. The framework does not replace Branch-and-CutAn exact MIP algorithm that builds a search tree of LP relaxations, pruning branches proven suboptimal and adding cutting planes to tighten the LP bound.; it reduces the number of nodes the solver must explore by starting it close to a high-quality feasible solution.

Why it matters: This paper addresses the exact computational bottleneck that makes AFRY BID3’s multi-scenario unit commitment analysis expensive: each scenario requires a fresh MILP solve, and warm-starting from a predicted solution compresses solve time without sacrificing optimality guarantees. For practitioners in energy systems who use FICO Xpress, Gurobi, or HiGHS for unit commitment, the deployable pattern is to train a transformer on historical commitment schedules, use it to warm-start the commercial solver, and fix confident binary variables before the search begins — a three-line change to the solver interface that can cut wall-clock time by 30–70 % on large instances. The 20 % rate at which the ML-warmed solver beats the solver alone is also a signal that ML-predicted solutions are capturing structural regularities in generator behaviour that the solver can then exploit.

arXiv:2604.21891 · 23 Apr 2026

Research Paper Algorithms Cutting Planes arXiv · Mar 2025

Beyond Local Selection: Global Cut Selection for Enhanced Mixed-Integer Programming 🔗

Foundation: Modern Branch-and-CutAn exact MIP algorithm that builds a search tree of LP relaxations, pruning branches proven suboptimal and adding cutting planes to tighten the LP bound. solvers generate hundreds of candidate Cutting PlanesValid linear inequalities added to an LP relaxation to tighten it toward the integer hull, reducing the gap between the LP bound and the optimal MIP solution. at each node of the search tree and must decide which to add to the LP relaxation. Too few cuts and the LP bound stays loose, requiring more branching. Too many cuts and the LP at each node becomes large and slow to solve. The selection heuristic matters enormously for solver performance — commercial solvers like Gurobi and FICO Xpress invest heavily in proprietary cut-scoring rules. The standard approach scores each cut locally: it looks at the cut in isolation and asks whether it is violated by the current LP solution and whether it has high curvature. The limitation is that local scoring ignores interactions between cuts: two individually mediocre cuts can jointly achieve better bound tightening than one locally excellent cut.

Researchers posting to arXiv (arXiv:2503.15847) propose a global cut selection framework that scores candidate Cutting PlanesValid linear inequalities added to an LP relaxation to tighten it toward the integer hull, reducing the gap between the LP bound and the optimal MIP solution. jointly rather than independently. The core insight is that cut selection is itself a combinatorial optimisation problem — choose a subset of cuts that maximises LP bound improvement — and that the local heuristic approximates it poorly when cuts have correlated geometry. The paper introduces a global scoring objective that accounts for cut-cut interactions and proposes an efficient greedy algorithm to maximise it. On standard MIP benchmark instances, global selection reduces the number of branch-and-bound nodes by up to 18 % compared to the local scoring baseline, with consistent gains across problem classes including unit commitment, vehicle routing, and network design. The results are reproducible using open-source solvers (HiGHS, SCIP) and do not require modifying solver internals — the selection logic runs as a cut callback.

Why it matters: Today’s term is Cutting Planes, and this paper goes directly to the algorithmic question practitioners least often ask: not “should I use cutting planes?” (yes, always, they are on by default) but “which cutting planes should the solver add, and does the selection rule matter?” The answer is yes, it matters, and the improvement is achievable as a solver callback without waiting for commercial vendors to update their proprietary heuristics. For any practitioner running unit commitment, lending portfolio optimisation, or supply chain MIP jobs that take more than five minutes, the first diagnostic is to profile where the solver spends its time — if it is generating many cuts but the LP bound is still loose, a global cut-scoring callback is the right intervention. The paper ships an open-source implementation.

arXiv:2503.15847 · Mar 2025

Term of the Day

Cutting Planes

A cutting plane is a valid inequality that every integer-feasible solution satisfies but the current fractional LP solution violates. Adding it to the LP tightens the feasible region without cutting off any integer point. — Ralph Gomory, Integer Programming and the Theory of Groups (1958)

A cutting plane (or cut) is a linear inequality that is: (1) valid — satisfied by every integer-feasible solution to the original problem; and (2) violated by the current fractional solution to the LP relaxation. Adding a cutting plane to the LP eliminates the fractional solution and forces the solver to find a better lower bound without removing any integer-feasible point. The process of generating and adding cuts repeatedly is called the cutting plane method. In modern solvers, cuts are applied at every node of the branch-and-bound tree — this combination is called branch-and-cut.

A concrete example

Suppose you are solving a unit commitment problem and the LP relaxation returns a fractional solution: generator G is “on” at 0.6 (60% of the time) in hour 3, even though the generator must be either fully on or off. A Gomory–Chvátal cut might look like: x₃ ≤ 0 or x₃ ≥ 1 (a simple integrality cut), or more usefully: a minimum-up-time cut that says “if G is on in hour 3, it must stay on in hours 4 and 5 too” — a constraint that the LP relaxation is not enforcing but every valid schedule must satisfy.

The solver checks whether the fractional solution violates this minimum-up-time inequality. It does (x₃ = 0.6 but x₄ = 0.2 violates the coupled constraint). The solver adds the inequality to the LP, re-solves, and gets a new fractional solution that is feasible with respect to the new cut. This process repeats until the LP bound is tight enough that branching can close the gap.

Why practitioners misread this

Cutting planes do not cut off integer solutions. The defining property of a valid cut is that every integer-feasible point satisfies it. A cut only removes the fractional region between the LP relaxation polytope and the integer hull. If a cut ever removes a feasible integer solution, it is not a valid cut and the solver has a bug.

Cuts are not always worth adding. Adding a cut tightens the LP bound but also makes the LP larger and slower to solve at each node. Commercial solvers like FICO Xpress and Gurobi use scoring heuristics to decide which cuts are worth their computational cost. The global cut selection paper (arXiv:2503.15847) demonstrates that the standard local scoring heuristic underperforms by up to 18 % on node count because it ignores interactions between cuts.

Gomory cuts are not the only cuts. Gomory–Chvátal cuts are the classic family, but modern solvers generate dozens of cut families: knapsack cover cuts, flow cover cuts, clique cuts, lift-and-project cuts, mixed integer rounding (MIR) cuts, and problem-specific cuts like minimum-up/down-time cuts for unit commitment or sub-tour elimination cuts for routing. The choice of cut family depends on the problem structure.

Where this shows up in practice

Cutting planes are active in every production Mixed Integer Programme (MIP) solve: supply chain planning, unit commitment, loan pricing, crew scheduling, vehicle routing, network design. They are not a setting the user turns on; they are the default behaviour of every commercial solver. The user-facing manifestation is the “cutting plane phase” in the solver log — the section before branching begins where the solver adds cuts to the root LP. For AFRY BID3 and Erste Group, the cutting plane phase is what makes their problems solvable at planning speed: without it, the LP relaxation is far too loose to prune the branch-and-bound tree efficiently.

Daily Synthesis

The shared mechanism Both AFRY BID3 and Erste Group are solving problems whose LP relaxations are hopelessly loose without valid inequalities: unit commitment has exponentially many feasible generator schedules, and loan pricing has a continuous price space with integer-like combinatorial interactions between credit-risk tiers and regulatory capital constraints. In both cases the solver earns its fee by adding cutting planes that slice off the fractional LP solution and force the search toward integer-feasible points. The FICO Xpress engine does this transparently in both deployments — the user never writes a cutting plane; the solver generates Gomory cuts, knapsack covers, and flow cover inequalities automatically.
The 22% vs 25% gap Erste Group reports a 22 % profit lift. AFRY reports 25 % user growth (a platform adoption metric, not a direct profit figure). These are not comparable numbers, but they illustrate a useful distinction: in regulated industries (banking, energy), the published outcome metric is often adoption or process change — not profit — because P&L attribution is contested and regulators require conservative reporting. The 22 % figure from Erste Group is unusually specific for a financial services deployment and likely reflects the FICO audit trail that quantifies contribution at the transaction level — something most banks cannot do with heuristic pricing.
The connection to cutting planes The unit commitment warm-start paper (arXiv:2604.21891) reduces the search space that the branch-and-cut solver must explore. The global cut selection paper (arXiv:2503.15847) makes the cutting plane step itself more efficient. Together they address the two computational bottlenecks in production MIP: starting the search well and pruning the tree fast. Neither paper replaces the solver — both treat the commercial solver as the correct tool and contribute ML-guided enhancements that run as callbacks or warm-start inputs.