Forecasting Intelligence  Lab

Research

The Research Proposal

An applied quantitative study connecting forecast accuracy to downstream inventory performance.

Background & problem

Greater sophistication is not automatically greater value.

Forecasting is central to inventory planning. Organisations are offered increasingly sophisticated tools — from exponential smoothing and ARIMA to machine learning and generative AI. But a model that lowers statistical error may not lower inventory cost: a slightly over-forecast raises holding cost, a slightly under-forecast raises stockouts, and the right balance depends on demand characteristics, lead time and the relative cost of holding vs stockouts.

The central research question is whether additional forecasting sophistication — from classical statistics to machine learning and small language models — necessarily leads to better accuracy and better inventory decisions.

Research gap

The existing literature has examined classical forecasting, machine learning, forecast accuracy, inventory performance and LLM forecasting separately, and research on M5 has shown better accuracy does not always translate into better inventory outcomes. The contribution here is an integrated, reproducible framework that varies forecasting sophistication and cost structure across two independent retail datasets, delivered as a public interactive artifact.

Research questions

Eight questions.

  1. RQ1 How do classical statistical methods, machine learning and small language models compare in demand forecasting accuracy?
  2. RQ2 Does increasing forecasting sophistication improve forecast accuracy?
  3. RQ3 Does improved forecast accuracy translate into improved inventory performance?
  4. RQ4 Does the most statistically accurate method produce the lowest inventory cost?
  5. RQ5 How does method performance vary across demand characteristics?
  6. RQ6 How does the relative cost of holding vs stockouts influence the operational value of methods?
  7. RQ7 Can a small locally deployed language model provide useful forecasting or inventory decision support?
  8. RQ8 Do the findings replicate across the M5 and Corporación Favorita datasets?

Objectives

What the study sets out to do.

  • 01 Establish naïve and classical statistical forecasting benchmarks
  • 02 Compare exponential smoothing and ARIMA/SARIMA methods
  • 03 Evaluate XGBoost as a machine-learning forecaster
  • 04 Experimentally evaluate a small local language model
  • 05 Compare methods on MAE, RMSE, WAPE and MASE
  • 06 Translate forecasts into inventory decisions via simulation
  • 07 Compare holding, stockout, total cost and service performance
  • 08 Investigate how demand characteristics influence performance
  • 09 Examine whether the most accurate model is also the most operationally valuable
  • 10 Validate findings across two independent retail datasets
  • 11 Provide managerial recommendations based on operational requirements

Full proposal

The complete narrative — background of the study, literature structure, hypotheses, conceptual framework, datasets, models, validation, inventory policy, cost sensitivity, statistical testing, LLM protocol, scope, timeline and deliverables — is maintained in Research_Proposal.md (v2.0), the canonical source of truth for this website.

To be updated after the final experiment. A public paper link will be added when a hosted version exists.