LSLauri Saarinen

Lauri Saarinen

*This is an algorithmically generated author record based on publicly available information.

Current Organizations / Institutes

The organizations/institutes from the scholar's latest article


Lauri Saarinen has a total of 14 co-authors. Here are the top co-authors by number of shared publications:


Latest Publications

  • Open Access

    Bullwhip on a Secular Shift: Price Cycles in the AI Infrastructure Boom

    by Lauri Saarinen

    Since late 2022 the supply chain for AI compute has shown the signature of a bullwhip on a secular demand shift: a shortage in 2023–24 with lead times of up to a year and rental prices three to four times their later lows, an easing in 2025 in which GPU rental prices fell by two thirds or more, and a renewed tightening from 2026 as hyperscaler capital expenditure re-accelerated and server memory contract prices rose by about 270 percent. We build a deliberately small system-dynamics model of this chain (two inventory echelons, a fab capacity stock with an acquisition delay, a grid capacity stock with a multi-year connection delay, and one endogenous chip price) in which every parameter and mechanism are declared, and a balanced-growth path is an equilibrium. Calibrated to 10 stylised facts from 2023–26 with no calendar-indexed forcing, the model reproduces the two-wave sequence that has been observed: order amplification rises from a demand peak of 2.4× the no-shock path to 5.7× at the fab and 24.2× in capacity investment, the chip shadow price peaks at 4.1× its technology trend, and the chain catches up with demand in 2025.3. Three results follow. First, the persistence of the cycle is governed by the capacity acquisition delay, not by the price response: price-responsive investment shortens shortages without destabilising the chain, and price-elastic demand damps the cycle and is beneficial to the overall business outcomes. Second, the grid constraint changes who faces shortages: with chips ordered on demand, up to 46% of demand is installed but unpowered at a five-year connection delay; gating chip orders on power moves the shortage upstream. When growth stops, the grid itself overbuilds by 3.9×. Third, take-or-pay contracts barely change physical flows but reallocate the cost of idle capacity to buyers in proportion to the contracted share, which is the mechanism behind the recent claim that a semiconductor bust would be shallower upstream than the boom.

  • Open Access

    Misvaluing Time: Salvage Modeling and Lead-Time Decisions

    by Lauri Saarinen, Hao-Chun Chuang, Bo Li, Suzanne de Treville

    The assumption that the value of unsold goods at the end of a demand period is constant is widespread in inventory theory. It is common, for example, to assume that a fashion product can be liquidated at a constant salvage value if still in inventory when the demand period has ended. We show that, under general mathematical conditions, this modeling assumption will cause companies to underestimate both the value of lead-time reduction and the cost of lead time increases. Assuming a constant residual value in an environment where salvage values are allowed to depend on the amount of overage will systematically bias the value of lead time negatively.

  • Open Access

    From Next-word to Next-World Prediction: Mapping Foundation Model Modalities to Operations Management Tasks

    by Lauri Saarinen

    AI adoption in operations management is increasingly shaped by foundation models, yet deployments often default to large language models even when operational information is primarily temporal, visual, or physical. This modality-task mismatch creates avoidable cost and risk and can degrade decision quality. We develop an evidence-based modality-task fit framework for production and operations by combining a PRISMA-compatible review of 44 primary foundation-model sources (2018-2026) with a synthesis of production research, using a transparent coding protocol and source-quality sensitivity checks. We classify six foundationmodel modalities (language, time-series, vision, vision-language, world-model, and visionlanguage-action) and calibrate their operational readiness. The resulting mapping matrix links modalities to nine OM task families and distinguishes where each modality is a primary candidate versus a supporting layer. The evidence supports three near-term design patternssolver-grounded LLM interfaces for planning and optimisation, zero-shot time-series forecasting, and VLM-assisted quality inspection-while revealing two major gaps for OM research: limited production-research validation of time-series foundation models and minimal OM integration of world models and VLA robotics. We translate these gaps into eleven testable propositions with measurement templates to enable OM-native evaluation and field validation in future work.

  • Open Access

    Heterogeneous but Segmentable: A Data-Driven Approach to Modelling Long-Term Care Trajectories in Multiple Sclerosis

    by Märt Vesinurm, Laura Mäkitie, Paul Lillrank, Lauri Saarinen, Paulus Torkki, Sini M Laakso, Miika Koskinen

    Managing chronic diseases with unpredictable care demand creates significant operational challenges for healthcare systems. Mapping long-term care trajectories is crucial for improving resource allocation, anticipating service needs, and designing efficient care pathways. We used a data-driven approach to map six-year care trajectories for 962 newly diagnosed multiple sclerosis patients, identify utilization clusters, and determine predictors of high utilization. We analyzed Event logs of remote, outpatient, emergency, and inpatient contacts from one year pre- to five years post-diagnosis using K-means clustering to identify utilization clusters, logistic regression to identify predictors of high utilization, and process mining to model variation between care trajectories. We identified two distinctive utilization clusters: a high-utilization cluster (14.1 % of patients) with persistently elevated annual encounter volumes across all care settings and low-utilization cluster (85.9 % of patients) with lower and declining use. Median service costs were €18,736 vs. €6,052 in high- and low-utilization clusters, respectively. Two or more early relapses were the strongest predictor of high utilization (OR = 6.33, 95 % CI 3.49-11.50, p < 0.001), with number of planned early remote and outpatient care contacts being also associated with future service utilization (OR = 1.07, 95 % CI 1.04–1.10, p < 0.001). High-utilization trajectories were approximately three times longer (82.4 vs 25.9 events) and more variable (3.1 vs 2.4 unique events per patient). These utilization clusters and their distinct trajectories provide a pragmatic segmentation of multiple sclerosis patients to support early identification of high-utilization subgroups and more robust capacity planning in specialist care. Highlights: We tracked the care trajectories of 962 people with relapsing–remitting multiple sclerosis using a Finnish population-based specialist-care datalake covering both inpatient and outpatient neurology services. Patients fell into two distinct utilization clusters: a high-utilization cluster with frequent contacts across all care settings and a low-utilization cluster with lower and declining use. Two or more early relapses, and the number of early outpatient and remote contacts were strong predictors of a patient’s long-term affiliation in the high-utilization cluster. Segmented care trajectories showed that high-utilization patients followed longer, more varied, and acute-oriented care patterns and had much higher service encounter costs. These findings can help clinicians and managers identify potential high-utilization patients early, target resources more effectively, and plan for future healthcare demand.

  • Open Access

    Happy Hour: Markdown Pricing to Reduce Food Waste

    by Xabier Barriola, Jan Holmström, Lauri Loikkanen, Lauri Saarinen

    Product expiration in grocery stores is the main cause of food waste at the retail level. To stimulate sales and minimize waste, retailers are resorting to end-of-day markdowns. This approach is costly to operate and it is impeded by lacking accurate inventory information. Investment in technology for inventory tracking and data-driven automated markdowns can provide more efficient solutions but business cases require evidence of impact. In this paper, we analyze the implementation of an automated markdown pricing solution in terms of food waste, prices, and gross margins. First, we use regression analyses to estimate the difference between a manual markdown process and an automated markdown solution. Second, we perform a differences-in-differences estimation to compare waste rates between treated and control product categories. Our results show that automated markdown pricing can improve the triple bottom line, benefitting the planet, retailers, and consumers. First, we observe that the automated solution led to a reduction in the food waste rate. Second, we find that the automated solution selected products that led to an increase in gross margin compared to the manual solution. Third, we show that consumers paid lower prices because higher discounts were given. Our study allows us to analyze the challenges and opportunities that online and physical retailers face when they decide to adopt automated markdown pricing as a solution for reducing waste. We present a qualitative evaluation of different solutions that are available for retailers and we propose a path forward.


  • Technical note: Option‐based costing and the volatility portfolio

    by Suzanne de Treville, Kyle Cattani, Lauri Saarinen

    It has been clearly established that a cost premium for responsiveness may be justified for profitable time‐sensitive products, and that this cost premium may suffice to render production in a high‐cost environment competitive. Time‐insensitive products considered in isolation seldom justify a cost premium, leading many decision makers to conclude that their production does not belong in a high‐cost environment. This leads to a manufacturing‐location decision in which profitable and time‐sensitive products are produced in a high‐cost environment and time‐insensitive products are transferred to a low‐cost environment. Responsiveness, however, requires a capacity buffer that provides the option to meet a demand peak for a profitable, time‐sensitive product. Leftover capacity can then be ideally deployed to manufacture time‐insensitive products to stock. We propose that the cost of the capacity buffer be considered as an option cost and assigned to the time‐sensitive product: “option‐based costing”. We then demonstrate use of demand volatility to create a portfolio of products that are time sensitive and insensitive to generate profit and increase competitiveness. Option‐based costing combined with a volatility portfolio reveals opportunities to produce competitively in high‐cost environments that have typically been considered unfeasible.