Forecasting and inventory intelligence for operations teams. A decision layer on top of the planning data your team already uses, with no data science or IT lift.
Forecasts are evidence. Decisions must survive being wrong.
Abstract. Forecasting and decision-making are different problems. Forecasting processes a large, noisy information set and compresses it into a smaller predictive information set consisting of expected outcomes and estimates of remaining uncertainty. Decisions must then be formulated from that reduced information set. In real industrial settings, limited data, changing regimes, and imperfect signals mean forecast error remains material. The shape, tails, correlations, and stability of the error distribution may themselves also be uncertain. More sophisticated uncertainty estimates can enrich the predictive information set, but converting that uncertainty into economically useful decisions remains a separate and difficult problem. Therefore, my preference for robust optimization is not a rejection of forecasting or probability. It is a preference for discipline when the probability law itself is uncertain. Forecasting should extract as much signal as the available information supports, while the decision layer converts that evidence into actions that remain economically viable across plausible forecast error and operational constraints without requiring a fully specified probability distribution. In controlled empirical results across two organizations, robust policies delivered higher realized profit and service than certainty-equivalent alternatives after controlling for inventory capital, with the evidence pointing to better inventory placement rather than larger buffers
Forecasting should extract as much useful signal as possible from the available information. Better data, richer features, ensembles, meta learning, rolling validation, and calibrated prediction intervals can all improve the predictive information set available to a decision maker. MetaLearner’s forecasting work follows this principle by searching for signal, selecting models dynamically, and explicitly quantifying uncertainty rather than hiding it behind a point estimate [1][2][3][4].
But forecasting and decision-making answer different questions. A forecast estimates what may happen. A decision must determine what to do given that estimate when inventory is finite, replenishment takes time, cash committed today constrains tomorrow, warehouses differ, and management simultaneously cares about service levels, turnover, and liquidity. Enterprise decisions are therefore multifaceted and stateful. Planners remain central to the decision-making process by setting objectives, defining the relevant constraints and trade-offs, and monitoring whether the resulting decisions remain aligned with operational reality.
Those constraints interact through time: today’s action changes tomorrow’s inventory, cash position, and feasible decision set. In that setting, the forecast should be an input into the decision system, not the decision itself [5][6].
Prediction intervals provide useful information about the range of possible outcomes. But using that uncertainty within probabilistic decision models can require stronger assumptions. Stochastic optimization is powerful when you can estimate the underlying probability distribution with sufficient confidence. Yet that requirement is stronger than it first appears. In realistic multi-period systems, this may also require assumptions about correlations between products and locations, tail behavior, and whether those relationships remain stable over time. This creates a second level of uncertainty. We are uncertain about future demand, but we may also be uncertain about the probability distribution used to describe it. MetaLearner’s earlier forecasting work explicitly considered probabilistic forecasts, Monte Carlo simulation, and stochastic programming, while also emphasizing prediction bands and uncertainty rather than relying solely on point estimates [4].
Simulation does not remove this distinction. A large number of Monte Carlo draws can estimate the consequences of an assumed distribution with increasing numerical precision. They do not provide additional evidence that the assumed distribution accurately describes the future.
More simulation can sharpen our understanding of a model. It cannot manufacture new information about the world.
That distinction matters more as compute becomes abundant. The Signal vs. Compute work makes the analogous point in forecasting: more computation does not automatically create proportionally more useful information. The advantage comes from identifying where predictive signal actually resides and allocating computation toward the features, models and observations that contribute information rather than redundancy [1].
Once hypothesis generation, model fitting and simulation become inexpensive, the scarce resource is increasingly independent information: new regimes, instruments, venues, supplier behavior, execution conditions and realized outcomes that the research process has not already optimized against. Repeatedly sampling from the same estimated historical distribution cannot substitute for genuinely new observations.
The same principle should govern optimization. The objective is not to manufacture the largest possible number of futures computationally. It is to identify which deviations from the forecast are sufficiently plausible and economically consequential that the decision should remain viable when they occur.
This leads to the question I care about most: What decision remains economically viable if the forecast is wrong in the plausible ways the evidence says it can be wrong? MetaLearner’s streamlined robust optimization framework uses observed forecast errors to calibrate the magnitude and asymmetry of plausible deviations around predicted demand, then jointly optimizes procurement and fulfillment subject to inventory, liquidity, lead time, and service constraints [5][6]. Rather than assigning precise probabilities to every possible outcome, the framework seeks decisions that remain satisfactory as uncertainty increases. This follows the robust satisficing approach developed in Streamlining Robustness, which uses data-driven robustness parameters and tractable formulations for sequential decision problems [6].
This gives forecast uncertainty a second purpose. It is not only a measure of confidence in a prediction. When uncertainty differs across products, locations, and time, it can help determine where to deploy scarce capital. Two products with the same expected demand can have very different uncertainty profiles, while forecast errors of the same magnitude can have different economic consequences because of margins, inventory positions, lead times, and liquidity constraints.
Robust optimization does not protect against every imaginable future. Its robustness parameters remain modeling choices and should evolve as new evidence arrives. The advantage is that these assumptions are explicit and inspectable rather than embedded in a fully specified probability law.
A useful decision system should not require reality to match the plan exactly. It should degrade gracefully as reality diverges from the forecast. This is where decision quality becomes distinguishable from prediction quality. A policy can be built on a reasonably accurate forecast yet remain highly fragile to the residual error, while another can start from the same forecast but translate that uncertainty into actions whose realized economic outcomes deteriorate more slowly.
We see this in an empirical study across two organizations. In both organizations, the robust policy had the smallest gap between planned and realized service. At the lowest tested uncertainty setting, that gap was 6.35 percentage points in Org1 versus 13.41 to 14.75 for the alternatives, and 4.10 percentage points in Org2 versus 8.72 to 9.35. Its capital efficiency also deteriorated less severely under the highest tested uncertainty, retaining 68.0% of baseline ROIC in Org1 versus 49.7% for the strongest alternative, and 37.0% in Org2 versus 31.4.
This aligns with the robust satisficing principle: the objective is not to eliminate forecast error, but to preserve acceptable performance as uncertainty increases [6].
None of this makes stochastic optimization inherently inferior. When distributions are stable, data are abundant, and probabilities can be estimated credibly, stochastic methods can exploit information that a conservative robust formulation may leave unused. MetaLearner’s earlier work explicitly recognized the value of Monte Carlo and stochastic programming in appropriate settings [4].
My preference is conditional on a different problem: distributional misspecification. When the cost of getting the probability law wrong is high, particularly in enterprise systems exposed to regime shifts, sparse tail observations, changing demand, operational constraints, and real liquidity consequences, I would rather be approximately right about the range of uncertainty and make a decision that survives it than be extremely precise about a distribution that may itself be wrong [5][6].
The philosophy is the same across forecasting and decision-making: do not confuse more computation with more information. Use compute aggressively to discover signal, improve forecasts and quantify uncertainty [1][2][3][4]. Then, when the forecast reaches the decision layer, remain disciplined about what the evidence actually supports [5][6]. The forecast should extract as much signal as possible from the information we have. The optimizer should remain disciplined about what we do not know.
Prediction should reduce uncertainty. It should not require us to pretend uncertainty has disappeared.
The forecast is evidence. The robust decision is the product.
[1] MetaLearner, “Signal vs. Compute.” https://www.metalearner.ai/blog/signal-compute
[2] MetaLearner, “MetaLearner Forecast.” https://www.metalearner.ai/blog/metalearner-forecast
[3] MetaLearner, “Next-Generation Forecast Pipeline.” https://www.metalearner.ai/blog/next-gen-forecast-pipeline
[4] MetaLearner, “Forecasting at Scale,” White Paper, October 2024. https://www.metalearner.ai/blog/blog-posts/white-paper/Forecasting_At_Scale_Oct24.pdf
[5] MetaLearner, “Robust Optimization.” https://www.metalearner.ai/blog/robust-optimization
[6] Chen, Z., Cheng, C., Chua, Y. J., Sim, M., Xiong, & P, Andikko, A. (2026). Streamlining robustness for sequential decision making. Working paper.
