TraceLink and Kinaxis plug trading-partner execution into concurrent planning; Mass General and GE HealthCare report 96 percent accuracy predicting operating-room missed-opportunity risk; and the Term of the Day is Conditional Value at Risk, the coherent risk measure that averages losses past a chosen quantile.
Transcript ▼
This is the Decision Optimisation Radar for 21 April 2026. Today’s thread: two production deployments where a risk or visibility signal becomes a live input to an operations-research scheduling or planning engine, plus a risk measure that underpins most modern stochastic optimisation.
First, TraceLink and Kinaxis announced an expanded strategic partnership delivering Kinaxis Network Collaboration by TraceLink. The integration embeds TraceLink’s Opus multienterprise network directly into Kinaxis Maestro, so planners see live, consented purchase-order and Advance Shipment Notice state from trading partners alongside their own concurrent-planning models. That matters because multi-tier visibility has been the binding constraint on agile replanning in regulated industries such as life sciences and consumer packaged goods.
Second, Surgical Solutions’ 2026 Perspective (April 2026) documents a Mass General and GE HealthCare pilot reporting 96 percent accuracy at flagging operating-room cases at high risk of cancellation or slot slippage. The risk score feeds an active reallocation workflow that drives a room-assignment optimiser, returning updated theatre plans to charge nurses on a sub-hourly cadence. Operating-room utilisation is one of the most studied scheduling problems in operations research, and a 96 percent precision target changes the expected-value calculation for live reallocation.
On the research side, arXiv 2604.07216 introduces an inexact trust-region method for structured nonsmooth optimisation and applies it to risk-averse stochastic programmes with Conditional Value at Risk and mean-deviation objectives. The authors report a two-to-five times speedup on two-stage portfolio and unit-commitment benchmarks with global convergence under inexact subproblem oracles. A second paper, arXiv 2604.12440, tackles Earth-observation satellite scheduling when operational rules are partially unknown to the solver. An active constraint-acquisition loop proposes candidate schedules, observes operator accept-or-reject feedback, and learns a constraint network that closes the gap with expert solutions in fewer than twenty queries, reporting a fourteen percent revisit-rate improvement over a static-rule Constraint Programming baseline. That paper is today’s Case Study.
Today’s term is Conditional Value at Risk. Conditional Value at Risk is the average loss you would suffer in the worst alpha fraction of outcomes. Unlike Value at Risk, which reports the threshold at that quantile and ignores the shape of the tail behind it, Conditional Value at Risk is coherent in the Artzner sense, admits a linear-programming reformulation via the Rockafellar-Uryasev identity, and penalises tail-thick distributions the way a risk manager actually experiences them. One-line version: Conditional Value at Risk is the mean of the worst slice of outcomes, not the boundary of that slice. That’s the Radar.
Industry Signals
TraceLink and Kinaxis Ship Network Collaboration: Live Trading-Partner Execution Feeds Maestro Concurrent Planning 🔗
Foundation: A contract-manufacturing network spins up a new vaccine product across half a dozen tier-one suppliers and twice that many tier-two component makers, and the planner at headquarters has to decide when to commit, when to hold, and when to reroute. Today those decisions run off a reconciled snapshot of purchase-order and shipment data that is two to five days old because every trading-partner system posts updates on its own cadence. Closing that lag requires a way to read live, consented execution state from trading partners directly into the planning engine, which is the pattern industry has been calling multi-tier execution visibility or multienterprise network collaboration.
TraceLink and Kinaxis announced on 15 April 2026 an expanded strategic partnership delivering Kinaxis Network Collaboration by TraceLink. The integration embeds TraceLink’s Opus multienterprise network directly into Kinaxis Maestro, so planners get live, consented visibility into trading-partner purchase-order and Advance Shipment Notice (ASN) state. Concurrent-planning scenarios can now run against real network signals instead of reconciled snapshots, collapsing the feedback loop between plan and execution for regulated industries where multi-tier visibility has been the binding constraint on agile replanning.
Mass General and GE HealthCare Report 96% Accuracy Predicting Operating-Room Missed-Opportunity Risk 🔗
Foundation: Picture a Monday morning in a tertiary hospital with eighteen operating rooms booked, three anaesthetists out, and a cardiac case that ran ninety minutes long on Sunday and pushed two elective knees into tomorrow’s slate. Every slot held by a case that will not start, and every slot freed by a cancellation, is a lost minute of theatre time that surgical teams describe as a missed opportunity. Predicting which cases will slip before they slip is a classic stochastic scheduling problem with human-in-the-loop rescheduling levers, and the central question is whether the predictor’s precision is high enough for a reallocation decision to be worth making.
Surgical Solutions’ 2026 Perspective (April 2026) documents preliminary results from a Mass General and GE HealthCare pilot of an artificial intelligence (AI) predicted missed-opportunity-risk model, reporting 96 percent accuracy at flagging cases at high risk of cancellation or slot slippage with enough lead time to re-sequence the day. The risk score drives an active reallocation workflow that feeds a room-assignment optimiser, returning updated theatre plans to charge nurses on a sub-hourly cadence rather than a daily one.
Research Papers
Inexact Trust-Region Method Cuts Solve Time 2-5x on Risk-Averse Stochastic Programmes with Conditional-Value-at-Risk Objectives 🔗
Foundation: A utility has to decide tonight which power plants to run tomorrow, before it knows what tomorrow’s demand will be. Planning for average demand is cheap but blind to risk — a handful of high-loss days can erase a year of thin margins. Planning for the absolute worst day is safe but wasteful. Conditional Value at Risk (CVaR) picks the middle: keep the average of the worst five percent of days within budget. The catch is that which days count as “the worst five percent” depends on the plan itself, so the underlying math problem has sharp edges — places where the best direction of improvement flips as one scenario crosses into or out of the tail. Off-the-shelf solvers stumble at those edges. Trust-region methods handle them by taking small, self-adjusting steps inside a bubble of confidence that shrinks whenever a step surprises them.
Submitted to arXiv on 9 April 2026, the paper (arXiv:2604.07216) introduces an inexact trust-region framework for structured nonsmooth optimisation and applies it to risk-averse Stochastic Programming (SP)Optimisation under uncertainty modelled by explicit probability distributions, typically expressed via scenarios and recourse. First explained 2 Apr 2026. with Conditional Value at Risk and mean-deviation objectives. The authors prove global convergence when each subgradient oracle returns only an approximate answer and report a two-to-five times speedup on two-stage stochastic portfolio and unit-commitment benchmarks versus sample-average-cut baselines. The method is solver-agnostic and slots in wherever a subgradient is expensive to evaluate.
Active Constraint Acquisition Closes a 14% Revisit-Rate Gap on Earth-Observation Satellite Scheduling in Under 20 Operator Queries 🔗
Foundation: An Earth-observation satellite operator sketches a scheduling problem in clear terms (visit these ground targets, respect the battery and thermal limits, minimise revisit time) and hands it to a solver team, which writes a Constraint Programming (CP)A paradigm that searches feasible assignments using propagation and global constraints rather than linear-programming relaxations. First explained 18 Apr 2026. model and returns a schedule. The operator then rejects the schedule because it violates rules the solver was never told about: a coastal imaging pass must precede a cloud-prone inland pass in the same orbit, or a target adjacent to a protected zone must be scheduled at dawn. Those rules exist in the operator’s head, not in the formulation, and they are the last fifteen percent of every real scheduling problem. Active constraint acquisition is the pattern of learning them from accept-or-reject feedback instead of asking the operator to enumerate them up front.
Submitted to arXiv on 14 April 2026, the paper (arXiv:2604.12440) formulates the problem as a learning-augmented scheduling loop: an acquisition policy proposes a candidate schedule, the operator accepts or rejects it, and a constraint network is updated from the feedback before the next proposal. The authors report a fourteen percent revisit-rate improvement over a static-rule Constraint Programming baseline with fewer than twenty operator queries before convergence. The full protocol is walked through in the Case Study panel below.
arXiv:2604.12440 · 14 Apr 2026📋 Case Study below ↓Term of the Day
Conditional Value at Risk
Do not cross a river if it is four feet deep on average. — Nassim Nicholas Taleb, Incerto (2012)
Conditional Value at Risk (CVaR) is the average loss incurred across the worst alpha fraction of outcomes. It fixes a tail probability (say, five percent), identifies the loss threshold at that probability (which is Value at Risk, or VaR), and then averages the losses that fall beyond it. That averaging step is what distinguishes it from Value at Risk: Value at Risk reports the boundary of the tail; Conditional Value at Risk reports the mean of what lives inside the tail.
A concrete example
Imagine a two-stage inventory plan for a seasonal product with one hundred demand scenarios. The planner’s objective could be written as expected cost, average-cost-across-worst-five-scenarios, or maximum-cost-across-all-scenarios. Each encodes a different appetite for tail risk.
Expected cost treats the worst scenario (a complete stock-out costing one hundred thousand euros) the same as a mediocre scenario (costing ten thousand euros) weighted by probability. That can leave the plan exposed when the worst five percent of scenarios are close in probability mass and collectively cost six hundred thousand euros.
A Conditional Value at Risk objective at the five percent level identifies the five worst scenarios and averages their cost: six hundred thousand divided by five equals one hundred twenty thousand euros per worst-five scenario. The plan then minimises that average. Because the objective is a convex function of the scenario costs, the resulting problem is a linear programme once the Rockafellar-Uryasev identity is applied. A Value at Risk objective would have reported only the boundary of the tail (ninety-five thousand euros), ignoring how heavy the losses beyond the boundary actually are.
Why practitioners misread this
It is not Value at Risk. Value at Risk reports a quantile of the loss distribution, the threshold exceeded with probability alpha. Conditional Value at Risk reports the mean of the losses that exceed that threshold. A tail-thick distribution can have the same Value at Risk as a thin-tailed distribution but a much higher Conditional Value at Risk, because the expected loss beyond the threshold is heavier. Reporting only Value at Risk obscures exactly that distinction.
It is coherent; Value at Risk is not. A risk measure is coherent when it is monotonic, positively homogeneous, translation-invariant, and sub-additive. Sub-additivity is the property that diversification never makes the combined portfolio riskier than the sum of the parts. Value at Risk can violate sub-additivity, which is why two safe portfolios can combine into a risky one under a Value at Risk regime. Conditional Value at Risk satisfies all four axioms simultaneously, which is why it is the default risk measure in modern regulatory frameworks such as the Basel III Fundamental Review of the Trading Book.
The Rockafellar-Uryasev reformulation makes it a linear programme. The seminal result is that minimising Conditional Value at Risk at level alpha can be rewritten as minimising a specific convex function over the original decision variable plus one auxiliary variable representing the Value at Risk threshold. For discrete scenarios, the reformulation is linear and composes cleanly with two-stage stochastic programmes, chance-constrained programmes, and risk-averse Markov decision processes. That tractability is what lets it plug into the inexact trust-region method in today’s arXiv 2604.07216.
Where this shows up in practice
Conditional Value at Risk is the default risk measure in regulatory capital frameworks (Basel III Fundamental Review of the Trading Book replaces Value at Risk with Expected Shortfall, which is equivalent to Conditional Value at Risk for continuous distributions), in energy dispatch and unit-commitment optimisation, in supply-chain inventory positioning under tail-demand risk, in robust portfolio optimisation, and as the standard risk objective in risk-averse Markov Decision Process formulations of constrained reinforcement learning. The diagnostic question when you see a paper or product claim 'risk-averse optimisation' is whether the underlying risk measure is Conditional Value at Risk, Value at Risk, variance, or a coherent distortion measure: the choice controls whether diversification is penalised, whether the solver decomposes, and whether the result inherits the axioms that make the measure defensible to a risk committee.
- TraceLink+Kinaxis Embedding TraceLink’s Opus multienterprise network into Kinaxis Maestro closes the two-to-five-day lag between reconciled and live trading-partner execution, so concurrent-planning scenarios run against real purchase-order and Advance Shipment Notice (ASN) state in regulated industries where multi-tier visibility has been the binding constraint on agile replanning.
- MGH + GE HealthCare A Mass General and GE HealthCare pilot reports 96 percent accuracy at flagging operating-room (OR) cases at high risk of cancellation or slot slippage, feeding an active reallocation workflow that drives a room-assignment optimiser on a sub-hourly cadence and turns a research-grade risk score into a live input to daily block-scheduling.
- Trust-Region CVaR An inexact trust-region method for structured nonsmooth optimisation delivers a two-to-five times speedup on two-stage stochastic portfolio and unit-commitment benchmarks with Conditional Value at Risk objectives, because inner subproblem evaluation (not outer cut generation) is usually the binding bottleneck.
- Active Constraint Acquisition Wrapping a Constraint Programming scheduler with an active constraint-acquisition loop closes a fourteen percent revisit-rate gap versus a static-rule baseline on Earth-observation satellite benchmarks in fewer than twenty operator accept-or-reject queries, offering a template for any scheduling domain whose last fifteen percent of formulation lives in operator judgment.