Sixteen real decisions behind Volume II, told from the published record.
Every chapter of Volume II carries an In Practice case: a real organization that took a decision of the chapter's kind, told from its published record. Each case sets out the decision, the method, the result and what it took to make it work, then what Meridian, the book's case company, takes from it and four questions for a board.
The cards give a preview. The full two-page case is in the chapter, after its Worked Example and before its AXIOM Lab.
Every morning, UPS's 55,000 U.S. drivers set out on routes of typically about 140 to 160 stops each, and one mile per driver per day is worth about $50 million a year. ORION, the system that now puts every route in order, nearly failed in testing before it saved $320 million by December 2015 and won the 2016 Franz Edelman Award. Yet its answers come with no certificate of optimality. How does a board know how good a good answer is?
In 1999 Kellogg, then the world's largest cereal producer, sold nearly $7 billion worldwide and coordinated more than 600 cereal stock-keeping units across about 27 plants, co-packers and distribution centers. Its planning system, a large linear program run week by week and month by month, cut costs by an estimated $4.5 million in 1995 and showed that simply closing a little-used line would leave too little capacity. What does it take to declare a whole supply network to a model?
$2.1–3.0 billion saved in 2007–2010, by MISO's estimate
Franz Edelman Award, 2011
Every five minutes, MISO decides how much each power plant on its grid should produce, and the price of energy at every point of it; around the time of its award, that meant more than 1,500 plants and about 40 million people. Those prices are the shadow prices of an optimization. By MISO's estimate, its markets let the region save $2.1 to $3.0 billion from 2007 through 2010, and the work won the 2011 Franz Edelman Award. What happens when a model's multipliers become the market's prices?
More than 80,000 feasible schedules analyzed for 2021
272 games, 18 weeks, 576 possible game windows and, by the solver maker's count, options well into the trillions. Not long ago four people had ten weeks to schedule 256 games. For 2021 the NFL analyzed more than 80,000 feasible schedules; by 2023, 4,000 computers produced dozens of options each night. “So we've got 5,000 Vals out there,” said the NFL's Mike North, recalling the days when Val Pinchbeck slotted games one by one. But is the schedule the league picks the best one, and could anyone prove it?
Clearance revenue up about 6 percent in a controlled pilot
Zara's clearance sales brought in more than a billion euros in 2008. Until 2007, a committee of four or five executives set the markdowns by comparing each stock's selling pace with the time left. Then came a model that plans every remaining week, carries out only this week's prices, and plans again next week. In a controlled pilot in Belgium and Ireland, clearance revenue rose about 6 percent, and Zara made the model its standard markdown tool. Why re-plan a plan you just made?
Every prawn caught this season is one that will not grow and breed for the next. Australia's Northern Prawn Fishery is managed not for the largest catch but for the most valuable stream of profits over an indefinite horizon. Maximum economic yield became its target in 2003; today a bioeconomic model sets each year's fishing effort, aiming at a stock larger than the one that gives the biggest sustainable catch. When is the best use of an asset to leave more of it in place?
$1.4 billion over three years, by American's estimate
Franz Edelman Award, 1991
In 1977 American Airlines launched its Super Saver fares. Robert Crandall, who named the practice yield management, saw that most flights left with empty seats that cost almost nothing to fill. The catch was every flight's dilemma: a seat sold cheaply in advance may turn away a full-fare passenger later. American put the benefit at $1.4 billion over three years and won the 1991 Franz Edelman Award. The founder of one rival put it bluntly: “That was the end of our run.” What is the last seat worth?
Sell a large block of shares fast and your own selling pushes the price down; sell slowly and the market may move against you before you finish. Robert Almgren and Neil Chriss turned that dilemma into an optimal trading schedule. A Citigroup team, estimating the impact model from 29,509 of its own orders, reported in 2005 that it was being built into its execution software, and many brokers now offer algorithms that minimize implementation shortfall. How fast should you exit a position you cannot sell at once?
Water in a reservoir can be used only once. Release it today and you save fuel; keep it and you guard against future shortfalls. Brazil runs a large hydro-dominated power system, 162 hydro plants in its 2023 models, on a chain of stochastic optimization models that put a value on stored water and set the official spot price. Their predecessor's authors reported savings of about $260 million in 1979–84, and after the 2001 rationing risk aversion was built in. What is a unit of reserve worth?
Ford bought from 1,400 Tier 1 suppliers at 4,400 sites, and its managers reckoned that a traditional risk analysis of them would take two or three years and be obsolete on arrival. So Ford, with a team led by MIT's David Simchi-Levi, asked a question with no probabilities at all: if this site went down, how long would recovery take, and what would it cost? Some of the largest exposures sat in unlikely places, such as a low-cost sensor that procurement was paying little attention to. Where is your risk hiding?
Futures with Lake Mead below 1,000 feet cut from about 19 to 3 percent
Nearly 40 million Americans rely on the Colorado River for drinking water. Planning to 2060, the U.S. Bureau of Reclamation and the seven Basin States faced four descriptions of the river's future: a century of records, reconstructions spanning nearly 1,250 years, a blend of the two, and 112 climate projections. They relied on none of them alone. Against the observed record, the climate projections raised the 2060 annual bill by $1.1 billion at the low end and $2.3 billion at the high end. How do you plan when you cannot know the odds?
Dutch Railways Weigh More Trains Against Punctuality
A record 87 percent punctuality in 2007
Franz Edelman Award, 2008
In December 2006 Netherlands Railways replaced a national timetable last overhauled in 1970, on a heavily used network where the alternative, doubling about 40 kilometers of track, would have cost about a billion euros. More trains or better punctuality? NS's board picked two of about ten radically different timetables, scored on punctuality, attractiveness to passengers and cost, and asked for one combining the best of both. Punctuality reached a record 87 percent in 2007, and the work won the 2008 Franz Edelman Award. Who should choose on a frontier?
A 10.06 percent improvement, after almost three years
In 2006 Netflix offered $1 million to anyone who could make its rating predictions 10 percent more accurate. Almost three years and more than 40,000 teams later, two coalitions finished in a dead heat, tied at the four decimal places the rules used. The winning blend of hundreds of models never went into production: the extra accuracy did not justify the engineering. Netflix still called the prize “a great return on investment.” When is a better prediction worth paying for?
About 30 percent cooling savings under direct control, by Google's account
Google had already cut its data centres' energy overhead sharply when DeepMind's neural networks, trained on sensor records, began recommending settings for a data centre's cooling plant; by Google's account, that plant used 40 percent less energy for cooling. Then Google handed the controls to the agent, inside limits set by its operators, behind a confidence filter, two layers of verification and a human override. Under direct control, Google reported, savings grew from 12 percent at launch to about 30 percent on average. How far should a learning agent be trusted to act on its own?
75 percent of potential in-flight events mitigated in 2008, by its account
Under its TotalCare agreements Rolls-Royce is paid by the flying hour, “so we are only rewarded for engines that perform,” and the call on each engine is its own: inspect it, send it to the shop, or let it fly on. Snapshots of 20 to 30 parameters, trended for small changes, reach analysts who confirm a problem before the airline is told. By the company's account, monitoring mitigated 75 percent of potential in-flight events in 2008; it did not prevent the Trent 1000's durability costs. What can a twin see, and what can't it?
Before the global financial crisis, some banks came to view funding liquidity as essentially free and its risk as essentially zero, and their business lines behaved accordingly. UBS funded every line of business, proprietary trading included, on the same, relatively favorable terms. When the crisis struck, supervisors found that the firms that fared better had priced liquidity internally, charging their lines for the contingent liquidity exposures they built. Funds transfer pricing became a supervisory expectation. What happens when a shared resource has no internal price?
Company and organization names identify the organizations whose published decisions the cases discuss. They are trademarks of their respective owners; no affiliation with or endorsement by them is implied.
