Forecasting and Optimization technologies

Over the last decade, the practice of data-driven decision-making in supply chains has evolved dramatically. Lokad began in 2008 with a focus on accurate forecasting, but the modern supply chain can’t afford to stop at mere predictions. Instead, decisions must be optimized under uncertainty. Lokad’s approach unifies forecasting and optimization into a single pipeline powered by cloud-based computing, programmatic paradigms, and a commitment to real-world performance.

In 2020, Lokad ranked No1 world-wide at the SKU level in the prestigious M5 forecasting competition, illustrating our relentless focus on accuracy. Yet accuracy alone isn’t enough: we must turn forecasts into decisions in the presence of tight constraints, volatile demand, and economic trade-offs. Lokad addresses these demands through probabilistic and stochastic approaches integrated into Envision, our domain-specific language.

1. Lokad’s Technological Generations

Lokad’s technology didn’t happen overnight; it evolved through multiple generations, each one addressing new challenges in supply chain analytics.

• Latent Optimization (2024)
  A paradigm designed to tackle hard, complex combinatorial scheduling and resource-allocation problems under uncertainty.

• Stochastic Discrete Descent (2021)
  A robust way to compute decisions when uncertainty dominates, using powerful stochastic optimization techniques.

• Differentiable Programming (2019)
  The convergence of numerical optimization and machine learning, delivering unified models that address real-world supply chain constraints.

• Deep Learning (2018)
  Leveraging AI-powered forecasts at large scale—this marked a shift from classical statistical methods to GPU-accelerated techniques.

• Probabilistic Forecasting (2016)
  An explicit emphasis on estimating full probability distributions of demand rather than single-point estimates.

• Quantile Grids (2015)
  Addressing supply chain constraints by computing entire distributions, not just average or median demands.

• Quantile Forecasts (2012)
  A move away from pure mean forecasts by introducing asymmetric “biased” forecasts aligned with business economics.

• Classic Forecasts (2008)
  Our original approach, benchmarked internally across a library of models, though now superseded by more sophisticated paradigms.

2. Beyond Forecasting: Why Optimization Matters

Classic forecasting provides a single numeric estimate—often a median—of future demand. While it’s useful for intuition, it leaves a critical gap for actual decision-making. Supply chains must deal with:

• Inventory constraints: Stock levels, supplier MOQs, lead times, etc.
• Economic trade-offs: Holding costs, shortage penalties, and obsolescence risks.
• Complex flows: Multi-echelon networks, uncertain lead times, multi-sourcing.

Lokad’s newest developments, such as Stochastic Discrete Descent and Latent Optimization, address these challenges by seamlessly weaving uncertainty into decision workflows—an approach that goes far beyond a mere “forecasting engine.”

3. How Lokad Operates in Practice

Our team of Supply Chain Scientists spearhead the initiative, handling the technical contributions, most notably all the Envision programming.

Step 1. Data Integration

We ingest historical transactions, product attributes, supplier information, and more. This unified dataset is the bedrock for both forecasting and optimization.

Step 2. Probabilistic Modeling

Instead of returning a single point forecast, Lokad’s methods estimate probabilities across multiple outcomes—useful for slow-moving SKUs or spiky demand. This embrace of uncertainty is key to robust planning.

Step 3. Decision Optimization

Through paradigms like latent optimization or stochastic discrete descent, we produce actual decisions—optimal reorder quantities, production schedules, or transfers—customized to your constraints and objectives.

Step 4. Continuous Improvement

As new data arrives, models are recalibrated quickly, and the decisions adapt automatically. This end-to-end loop ensures supply chain practitioners stay agile and outmaneuver shifts in demand or supply.

4. Envision and White-Boxing

A Domain-Specific Language for Supply Chain

Lokad does not hide its technology behind an opaque, “one-size-fits-all” engine. Instead, we provide Envision, a language designed for transparent and configurable supply chain analytics. Every step of the pipeline can be inspected and adapted.

Tailoring to Business Realities

Because supply chains differ widely—production vs. retail vs. MRO—Envision scripts let your teams, and our Supply Chain Scientists, hard-code constraints or heuristics specific to your processes. Coupled with Lokad’s advanced predictive capabilities, this white-box approach solves your actual problems rather than fitting you to a rigid template.

5. Next Steps

Lokad started in 2008 with a straightforward promise: accurate forecasts. We now merge those forecasts with robust optimization to deliver superior decision-making under uncertainty. Whether you struggle with tight scheduling, spiky demand, or multi-echelon flows, Lokad’s generational technology—from quantile forecasting to latent optimization—has you covered.

Curious to learn more? We invite you to:

• Dive into Latent Optimization if you’re facing hard, combinatorial scheduling challenges.
• Explore Stochastic Discrete Descent if you want to see uncertainty integrated into your day-to-day decisions.
• Check out Differentiable Programming for a deeper look at modern machine learning fused with supply chain optimization.
• Or get in touch for a personalized demo to see how Lokad can precisely model your business constraints.

Ultimately, forecasting and optimization go hand in hand—Lokad’s role is to ensure you benefit from the best of both worlds.