Intro & scope
Forecasting and risk-scoring use cases (EU/GDPR focus) in four patterns: DEMAND (volumes, prices, load, cash), CUSTOMER (churn, CLV, leads, returns, no-shows, attrition), ASSET (machines, batteries, networks, weather, sensors, quality) and RISK (credit, insurance, claims, audit, public-sector and valuation risk). Mostly tabular and time-series ML (gradient boosting, statistical and pretrained time-series models) with an LLM only for explanation and workflow. Scope boundaries: PRED-001 is unit demand, PRED-011 staffing/workload, PRED-016 cash flow, PRED-007 energy load and PRED-008 prices; PRED-006 is machine failure, PRED-014 process quality/soft sensors, PRED-025 field failures and PRED-033 buried networks; PRED-009 is retail/corporate credit, PRED-021 payment behaviour in collections and PRED-010/030/037 insurance; PRED-017 is weather, PRED-027 floods, PRED-036 climate-risk and PRED-042 air quality; PRED-028, 029, 039 and 045 are public-sector risk profiling. Individual-level scoring (PRED-009, 012, 029, 045 and often 002/010/021/022/030/043) is high-risk or borderline under the AI Act and needs DPIA, human decision, bias testing and (for HR) works-council involvement; Art. 5 bans apply to profiling-only crime prediction. Cross-family: OPT (act on forecasts), ANOM (detect rather than predict), DATA (feature pipelines), VIS (dashboards).
Use cases
| ID | Title | Pattern | Maturity | Path | HW tier | AI Act risk |
|---|---|---|---|---|---|---|
| PRED-001 | Demand forecasting (retail, manufacturing, wholesale) | PRED.DEMAND | mainstream | A | HW1 | minimal |
| PRED-002 | Sales forecasting & opportunity win prediction | PRED.CUSTOMER | mainstream | A | HW1 | depends |
| PRED-003 | Customer churn prediction | PRED.CUSTOMER | mainstream | A | HW1 | minimal |
| PRED-004 | Customer lifetime value prediction | PRED.CUSTOMER | mainstream | A | HW1 | minimal |
| PRED-005 | Lead scoring | PRED.CUSTOMER | mainstream | A | HW1 | minimal |
| PRED-006 | Predictive maintenance of machines & equipment | PRED.ASSET | mainstream | A | HW1 | minimal |
| PRED-007 | Energy load & renewable generation forecasting | PRED.DEMAND | mainstream | A | HW1 | minimal |
| PRED-008 | Price forecasting (commodities, energy, raw materials) | PRED.DEMAND | mainstream | A | HW1 | minimal |
| PRED-009 | Credit scoring & default prediction | PRED.RISK | mainstream | A | HW1 | high |
| PRED-010 | Insurance risk pricing & underwriting | PRED.RISK | mainstream | A | HW1 | depends |
| PRED-011 | Workload & staffing demand forecasting (call volume, patient volume) | PRED.DEMAND | mainstream | A | HW1 | minimal |
| PRED-012 | Employee attrition prediction | PRED.CUSTOMER | early-adopter | A | HW1 | high |
| PRED-013 | Yield prediction (agriculture & production processes) | PRED.ASSET | early-adopter | A | HW1 | minimal |
| PRED-014 | Process quality prediction & virtual sensors (soft sensors) | PRED.ASSET | mainstream | A | HW1 | depends |
| PRED-015 | Automated real-estate valuation (AVM) | PRED.RISK | mainstream | A | HW1 | depends |
| PRED-016 | Cash-flow & liquidity forecasting | PRED.DEMAND | mainstream | A | HW1 | minimal |
| PRED-017 | AI weather forecasting & weather impact prediction | PRED.ASSET | early-adopter | A | HW3 | depends |
| PRED-018 | ETA & delivery-time prediction | PRED.ASSET | mainstream | A | HW1 | minimal |
| PRED-019 | Traffic & mobility demand forecasting | PRED.DEMAND | mainstream | A | HW1 | minimal |
| PRED-020 | E-commerce return prediction & return-rate reduction | PRED.CUSTOMER | early-adopter | A | HW1 | minimal |
| PRED-021 | Payment-behaviour prediction & collections prioritisation (AR, dunning, debt recovery) | PRED.RISK | mainstream | A | HW1 | depends |
| PRED-022 | Appointment no-show prediction & smart overbooking | PRED.CUSTOMER | early-adopter | A | HW1 | depends |
| PRED-023 | Project schedule & cost-overrun risk forecasting | PRED.RISK | early-adopter | A | HW1 | depends |
| PRED-024 | Predictive process monitoring (case delay & outcome prediction from event logs) | PRED.ASSET | early-adopter | A | HW1 | depends |
| PRED-025 | Warranty claims forecasting & field-failure early warning | PRED.ASSET | early-adopter | B | HW1 | minimal |
| PRED-026 | Battery state-of-health & remaining-useful-life prediction (EV, stationary storage, battery passport) | PRED.ASSET | early-adopter | A | HW1 | minimal |
| PRED-027 | Flood & hydrological forecasting and early warning | PRED.ASSET | early-adopter | A | HW2 | depends |
| PRED-028 | Risk-based audit & inspection case selection (tax, customs, food safety, labour inspection) | PRED.RISK | early-adopter | C | HW1 | depends |
| PRED-029 | Public-service client risk profiling (jobseeker integration, social-welfare triage) | PRED.RISK | early-adopter | D | HW1 | high |
| PRED-030 | Telematics driving-behaviour scoring & usage-based motor insurance | PRED.RISK | mainstream | A | HW1 | depends |
| PRED-031 | Smart-meter energy disaggregation (NILM) & personalised energy-saving insights | PRED.ASSET | early-adopter | A | HW1 | minimal |
| PRED-032 | Manufacturing cost estimation & instant quoting from CAD (should-cost) | PRED.ASSET | mainstream | A | HW1 | minimal |
| PRED-033 | Buried network failure prediction & renewal planning (water mains, gas pipes, cables) | PRED.ASSET | early-adopter | A | HW1 | depends |
| PRED-034 | Visual trend forecasting for fashion & consumer products | PRED.DEMAND | early-adopter | A | HW2 | minimal |
| PRED-035 | Securities settlement-fail prediction & post-trade exception management | PRED.RISK | early-adopter | B | HW1 | minimal |
| PRED-036 | Physical climate & natural-catastrophe risk assessment for assets & portfolios | PRED.RISK | mainstream | A | HW2 | depends |
| PRED-037 | Claims severity, litigation propensity & loss-reserving prediction | PRED.RISK | mainstream | A | HW1 | depends |
| PRED-038 | Litigation outcome prediction & legal analytics | PRED.RISK | early-adopter | A | HW2 | depends |
| PRED-039 | Place-based crime hotspot forecasting | PRED.RISK | early-adopter | A | HW1 | depends |
| PRED-040 | Humanitarian crisis & displacement forecasting for anticipatory action | PRED.RISK | early-adopter | B | HW1 | depends |
| PRED-041 | Used-vehicle valuation & residual-value forecasting (leasing, fleets, remarketing) | PRED.RISK | mainstream | A | HW1 | minimal |
| PRED-042 | Hyperlocal air-quality forecasting & pollution source attribution | PRED.ASSET | early-adopter | A | HW2 | minimal |
| PRED-043 | Athlete injury-risk prediction & training-load management | PRED.RISK | early-adopter | A | HW1 | depends |
| PRED-044 | Sensor-based livestock health, heat & calving prediction (boluses, collars, ear tags) | PRED.ASSET | mainstream | A | HW1 | minimal |
| PRED-045 | Recidivism & offender risk assessment in criminal justice (probation, prisons) | PRED.RISK | mainstream | A | HW1 | high |
| PRED-046 | Structural health monitoring of bridges & buildings from sensor data | PRED.ASSET | early-adopter | A | HW1 | depends |
| PRED-047 | Economic nowcasting & macro/tax-revenue forecasting (central banks, ministries, institutes) | PRED.DEMAND | early-adopter | C | HW1 | minimal |
| PRED-048 | Building energy performance prediction & renovation roadmap generation (EPC / iSFP) | PRED.ASSET | early-adopter | A | HW2 | minimal |
Shared stack
Common layers: object storage plus a warehouse (PostgreSQL/DuckDB, TimescaleDB or ClickHouse for series), dbt/Dagster or Airflow pipelines, a feature store (Feast or plain tables), MLflow for experiments and model registry, an evaluation layer (backtesting, calibration, prediction intervals, drift via Evidently), SHAP-style explanations, and a serving layer (batch jobs plus FastAPI/BentoML). Add Fairlearn-style fairness audits and audit logging for every person-level model. Output goes to BI (Metabase/Superset) or ERP/CRM via the workflow engine (n8n/Temporal). The same stack serves OPT, ANOM, DATA and VIS.
- HW1 (CPU/laptop, no GPU): LightGBM/XGBoost/CatBoost, statsforecast/Prophet/ETS, Darts or sktime, conformal intervals, pretrained time-series models (Chronos-bolt / TimesFM small / Moirai small) on CPU, Postgres/DuckDB, MLflow, a 3-8B quantised LLM (llama.cpp/Ollama) for narrative explanations. Covers demand, customer, credit, insurance, maintenance, cash-flow, SHM, livestock and most RISK cases.
- HW2 (single 24-48 GB GPU): all of HW1 plus fine-tuned or larger pretrained time-series models, graph neural nets, hydrological/air-quality surrogate models, VLM for images (fashion trends, property photos, CAD, vehicles), a 14-32B instruct model via vLLM, bge-m3 embeddings with Qdrant/pgvector for text-heavy cases (litigation, claims, warranty). For floods, air quality, climate-risk, building energy and visual trends.
- HW3 (multi-GPU 80 GB+ or EU-hosted API): global AI weather and Earth-observation foundation models (GraphCast/Pangu/Aurora class, ensembles), large-scale reprocessing of satellite data; the downstream impact layer still runs on HW1. Prefer EU weather-service data (ECMWF, DWD, Copernicus) and EU-hosted compute.
- Path guidance: A = managed/EU cloud forecasting or build on open-source ML; B = specialised vendor or integration with core systems (post-trade, warranty); C/D = public-sector or restricted deployments needing legal basis, audit and in-house control. Share one feature store, one model registry and one monitoring layer across use cases rather than per-model stacks; person-level models should sit in a separately governed zone with DPIA, access control and retention limits.