Price forecasting (commodities, energy, raw materials)

Since the 2021–2023 energy and raw-material shocks, the price of electricity, gas, steel, copper, polymers or wheat can decide whether a Mittelstand manufacturer or a Stadtwerk makes a profit. AI price forecasting combines market history, fundamental drivers (weather, renewables, fuel, CO2, inventories, FX) and, more and more, signals from text, and returns probabilistic forecasts and scenarios rather than a single number. Procurement, treasury and trading use them to time purchases, set hedge ratios, plan budgets and price their own quotes. The value comes from better decisions under uncertainty. Nobody, AI included, can reliably "predict the market".

Problem & value

• Who: buyers of energy and raw materials in manufacturing, food, chemicals, construction and logistics; utilities and Stadtwerke (portfolio and tariff pricing); energy traders and direct marketers; controlling and FP&A (budget prices); public procurement (energy tenders for municipalities).
• Today: buyers rely on broker newsletters, price reporting agency (PRA) reports, a futures curve and gut feeling. Budgets use one "plan price", and purchases are often timed by calendar rather than by market signal. Many SMBs have no structured view of their price risk at all.
• Why AI helps: price formation is non-linear and depends on many drivers (merit order, renewable infeed, fuel switching, inventory levels, policy). ML and time-series foundation models combine these drivers better than univariate rules and give calibrated uncertainty bands, which matter more for risk decisions than the point forecast does.
• Structural change in power: since delivery day 1 October 2025 the European single day-ahead coupling (SDAC) clears in 15-minute market time units instead of hourly ones [5]. That quadruples the number of day-ahead price points and makes intraday price shape (solar noon dips, evening peaks, negative prices) more important for anyone with flexible load, storage or PV.
• Benefit figures: vendors seldom publish measured, auditable savings from timing purchases. Count any claim as unverified until you have backtested it on your own purchase history (see KPIs).

Variants

• Short-term power prices (day-ahead / intraday, 15-min): for traders, direct marketers, storage and flexible industrial load. Fundamentals plus ML, horizon from hours to a few days. Feeds OPT-039 and OPT-011.
• Medium/long-term energy (gas, power, CO2 forward curves): for budgeting, procurement tranche strategies and tariff calculation at Stadtwerke. Horizon is months to years, and scenarios rather than point forecasts are the norm.
• Raw materials & metals for manufacturing: steel, aluminium, copper, nickel, polymers (PP, PE, PVC), paper/pulp. Usually LME/exchange prices or PRA indices plus surcharges. Horizon is 1–12 months, used for purchase timing, price-escalation clauses and quote calculation (see PRED-032).
• Agricultural & food commodities: wheat, corn, cocoa, coffee, sugar, dairy, vegetable oils. Weather and crop-condition data, satellite indices and trade flows are key drivers. Used by millers, food producers and agricultural traders.
• Freight & logistics rates: container, bulk and road-freight indices as an input cost.
• Public sector: municipal and state energy-purchasing pools that time tranche purchases; economic institutes producing commodity assumptions for macro forecasts (PRED-047).
• Private use: households on dynamic electricity tariffs (in Germany suppliers have had to offer them since 2025) shifting EV charging, heat-pump or battery use to cheap quarter-hours. In practice this is a feature of the tariff app or the home energy management system, not a stand-alone tool.

How it works

1. market data (exchange/PRA prices, futures curves) + fundamentals (load, wind/PV, fuels, CO2, FX, inventories) + weather forecasts + text signals (news, reports, policy) [optional]
2. data pipeline: alignment of time zones/MTUs, holidays, contract rolls, unit/currency normalisation
3. feature engineering (lags, calendar, residual load, spreads, curve shape)
4. model ensemble: statistical baseline (LEAR/ARIMA/ETS) + gradient boosting + deep/foundation TS model
5. probabilistic output (quantiles, scenarios) + calibration check
6. decision layer: buy/fix/wait rules, hedge ratios, budget scenarios, optimisation (OPT-*)
7. human review (buyer / risk committee / trader)
8. order, hedge or plan update
9. backtesting & monitoring against realised prices

• Models: for day-ahead power, strong regularised linear models (LEAR) and DNNs are the reproducible academic baselines from the open epftoolbox benchmark (EPEX-BE/FR/DE, Nord Pool, PJM) [3][4]. Gradient-boosted trees (LightGBM/XGBoost/CatBoost) on fundamental features are the workhorse in industry. Time-series foundation models (Chronos-2, TimesFM, Moirai) give strong zero-shot baselines and, since 2025/2026, accept covariates [1][2][7].
• LLMs do not produce the numeric forecast. They extract signals from news and reports (supply outages, policy changes, strikes), write the market commentary, and explain drivers to buyers. Treat LLM-generated price predictions as unreliable.
• Human-in-the-loop: forecasts are advisory. Purchase and hedge decisions stay with the buyer or the risk committee, within a documented procurement/hedging policy (limits, tranches, approval levels).

Data preparation

Needed are historical prices and indices (EPEX SPOT, EEX, LME, price reporting agencies), fundamental drivers (fuel, CO2, FX, weather) and the company's own purchase history and contracts. In an SMB the external data comes from broker mails, PDF market reports, subscription portals and Excel sheets, while the own purchases sit in the ERP and in contract files. Typical gaps: PDF-only reports, licence limits on redistributing price data, different units and currencies, missing contract terms in structured form, and prices copied by hand into spreadsheets.

Preparation steps: (1) choose the few price series that really drive costs, (2) subscribe to feeds or APIs where licences allow, (3) automate download and parsing of reports (n8n or Python scripts, OCR for PDFs), (4) normalise units, currencies and time zones in one table, (5) extract own purchase prices and volumes from the ERP, (6) keep a log of forecasts versus actuals for evaluation. Useful PREP cases: PREP-005, PREP-011, PREP-006. Also DATA-003 and DOC-018 for report extraction.

Rough share of effort: about 20-30 percent, since price series are mostly clean and numeric. Most of the work is licensing and automating the feeds. See GUIDE-003.

Implementation paths

Which path for whom

• SMB — Path A: Procurement and energy-price tools or broker services with ready forecasts and alerts cover the standard need at low cost.
• Enterprise — Path B: Forecasts should combine own contracts, volume plans and market feeds through APIs inside an existing procurement or risk system.
• Developer — Path C: Open-source time-series models on HW1 with licensed or public price data allow transparent, testable forecasts.
• Public — Path C: Public buyers and municipal utilities prefer self-operated or sovereign hosting and traceable methods for procurement and budgeting.
• Private — Path A: Tariff and fuel price apps give simple forecasts for private decisions.

Path A — Buy (off-the-shelf)

• Energy market forecasts & analytics: Volue (NO/EEA; SpotEx price forecasts, Insight) [8], Montel (NO/EEA), Aurora Energy Research (UK, with a Berlin office; long-term scenarios), energy & meteo systems/emsys (DE; renewables and price forecasts for traders), plus the forecasting services of energy suppliers and direct marketers. Check the market coverage (DE-LU zone, 15-min) and the delivery format (API, Excel).
• Raw-material & commodity forecasting for procurement: ChAI (UK; AI forecasts for metals, energy, plastics and agricultural commodities, plus a price-risk insurance product) [9]; Expana (formerly Mintec; food and agricultural commodity prices and forecasts); PRAs with forecast services (Argus, ICIS, Fastmarkets, S&P Global Commodity Insights). Some procurement suites (SAP Ariba/S/4HANA, Coupa, Jaggaer) integrate external price indices into category management.
• Choose A when: you are a buyer, not a trader. External forecasts plus a documented tranche strategy cover most SMB needs, and the vendor's data licences (exchange and PRA data) are often worth more than the model.

Path B — Build on cloud APIs

• Managed forecasting: Google BigQuery ML AI.FORECAST (TimesFM; also the licensed route to commercial use of TimesFM 3.0) [2]; Amazon SageMaker / Bedrock with Chronos models and AutoGluon-TimeSeries [1]; Azure Machine Learning AutoML forecasting; Nixtla TimeGPT API [10]. Pick EU regions (Frankfurt, Netherlands, Paris).
• Data: ENTSO-E Transparency Platform API (prices, load, generation forecasts for all EU zones), Bundesnetzagentur SMARD (German wholesale prices and generation, free download) [6], EEX DataSource (licensed) [11], exchange and PRA feeds, ECMWF/DWD weather.
• LLM layer: a frontier or EU model (e.g. Mistral, Claude, GPT, Gemini via EU endpoints) summarises news and reports into structured events (outage, tariff, sanction) that become features, and writes the weekly buyer briefing.
• Choose B when: you have a data team and specific price drivers (e.g. your own surcharge formulas, regional indices) that generic vendors do not cover.

Path C — Self-host (open source)

• Libraries: Nixtla statsforecast / neuralforecast (NHITS, TFT, PatchTST, TiDE and others, with exogenous support, Apache-2.0) [12]; Darts; sktime; AutoGluon-TimeSeries; LightGBM/XGBoost; epftoolbox for power-price baselines and evaluation [3].
• Open-weight foundation models: Chronos-2 (120M params, Apache-2.0, multivariate and covariate-aware, released Oct 2025) and Chronos-Bolt [1]; TimesFM 2.5 (Apache-2.0 weights). Note that TimesFM 3.0 weights (Aug 2026) are under a non-commercial licence [2]. Also Moirai 2.0 (Salesforce uni2ts; weights are CC-BY-NC-4.0, i.e. non-commercial) [7].
• Hardware: HW1 is enough. Gradient boosting and models with 9M–200M parameters run on CPU servers. A small GPU only speeds up fine-tuning or large backtests. Orchestration with Airflow/Dagster, storage in TimescaleDB/ClickHouse/Parquet, dashboards in Grafana/Superset.
• Ops effort: most of the work is data engineering (feeds, licences, contract rolls, 15-min/hourly mixing, daylight-saving shifts) and continuous backtesting, not the model.

Path D — Custom training

Worth it for traders, utilities and large buyers whose edge depends on proprietary fundamentals (own portfolio, plant availability, regional surcharges) or on very short horizons (intraday). Typical setups: fine-tune Chronos-2/TimesFM on several years of your own price and driver history, or train dedicated probabilistic models (quantile regression, distributional DNNs, conformal calibration). You need at least several years of clean history covering different regimes (pre-2021, crisis, post-crisis) and point-in-time copies of the forecasts that were actually available at decision time (to avoid look-ahead bias).

State of the art (2026)

• Foundation models went multivariate. Chronos-2 (Oct 2025) and TimesFM 3.0 (Aug 2026) natively handle past and known-future covariates [1][2]. Covariates matter most for price forecasting (residual load, fuel prices), and the Chronos authors report their largest gains on tasks with exogenous features [1]. Leaderboards such as GIFT-Eval and fev-bench rank these models, but they cover general time series and do not substitute for a backtest on your own market.
• Power-specific practice: hybrid fundamental + ML models dominate. Expert-designed features (residual load = load − wind − PV, gas/CO2-implied marginal cost) plus LEAR/GBM/DNN ensembles remain hard to beat. Evaluation follows the epftoolbox best practice: multi-year test periods, several markets, and statistical significance tests such as Diebold–Mariano [3][4]. The move to 15-minute MTUs (SDAC, Oct 2025) [5] forced most models to be retrained on a new resolution with only a short history.
• Probabilistic forecasting (quantiles, scenario sets, conformal prediction) is now expected. Buyers plan with P10/P50/P90 rather than one price.
• Text and LLM signals: LLM-extracted events from news, outage notices (REMIT inside-information platforms) and policy announcements are used as features or as analyst support. Evidence that they improve medium-term forecasts consistently is still mixed.
• Limits and failure modes: structural breaks (wars, sanctions, policy shocks, pandemics) are not in the training data. Long-horizon commodity forecasts often barely beat the futures curve. Models can look excellent in backtests through look-ahead leakage, so use point-in-time data. Negative and spike prices in power break naive error metrics. Forecasts that many market participants act on can become self-defeating.

Shared resources & combinations

• Same time-series stack (time-series + tabular-ml, HW1): PRED-001 demand forecasting, PRED-007 energy load & renewable generation forecasting (its outputs are core features for power prices), PRED-016 cash-flow forecasting (price scenarios drive cost lines), PRED-047 economic nowcasting. One feature store, one forecasting service (e.g. AutoGluon/Nixtla + Chronos-2) and one backtesting framework can serve all of them.
• Weather inputs: shared with PRED-017 AI weather forecasting and PRED-007.
• Downstream decisions: OPT-039 algorithmic power trading and OPT-011 battery storage dispatch (both consume 15-min price forecasts); OPT-004 dynamic pricing (pass-through of input costs); OPT-010 portfolio/hedging; AGENT-005 procurement agent (purchase timing, supplier negotiation with price outlook); PRED-032 should-cost / quote calculation (material cost component); DATA-024 FP&A copilot (budget price scenarios).
• Text signals: LANG-017 market intelligence and ANOM-011 supplier risk monitoring share the web-scraping + llm-general pipeline that turns news into structured events. DATA-015 spend analytics tells you which commodities actually matter for your spend.

Effort, cost & KPIs

• Path A: days to weeks. Cost is mainly subscription and data licences (exchange and PRA data can cost more than the forecasting software).
• Path B/C: 1–3 months for a first production pipeline on 3–10 key commodities, most of it data integration. Compute cost is negligible, but data licences and one data scientist/engineer are not.
• Path D: quarters, with a dedicated quant/data team. Only pays off at trading or large-portfolio scale.
• KPIs: forecast error vs. naive and futures-curve benchmarks (MAE, sMAPE, pinball loss/CRPS for quantiles, coverage of P10–P90 bands); procurement KPIs: realised purchase price vs. market average or benchmark index over the period, budget variance, hedge effectiveness; for power, value of flexibility captured (EUR/MWh). Always measure against a "buy evenly over time" baseline.

People & change

Involve management (risk appetite, hedging policy) and the head of purchasing or energy management as process owner. Buyers are the end users and must trust and challenge the forecast. IT sets up data feeds and access; legal and compliance check data licences, hedging rules and, where relevant, market regulation. Suppliers and brokers should be informed that forecasts complement, not replace, their advice.

Works council: no in the typical case, because the system forecasts market prices and does not monitor employees; if purchasing decisions are logged per person for evaluation, §87(1) no. 6 BetrVG could apply, so keep logs on the process level. §90(1) no. 3 still requires early information of an existing works council about AI-supported planning.

Data protection: price data is not personal; a DPO is only needed if contact data or personal notes are included. Support AI literacy (Art. 4 AI Act) with a session on forecast uncertainty: prices are partly unpredictable and a forecast is no guarantee (GUIDE-006). Pilot with one commodity and one buyer, and track forecast versus actual before using it for decisions (GUIDE-009). Typical concerns: false precision (show ranges and scenarios) and blame when the forecast is wrong (document that the decision remains with the buyer). Humans decide on every purchase and hedge. See GUIDE-002.

Risks, regulation & compliance

• EU AI Act: minimal risk in the typical form. Price forecasting is not an Annex III category and involves no Art. 5 practice; the Art. 4 AI literacy duty applies. Using price forecasts inside critical-infrastructure safety components would be a different case, but market forecasting is not safety control [13].
• GDPR: normally no personal data. The exception is household or SME smart-meter load data used for customer-specific tariff forecasts, which needs a legal basis and minimisation.
• Energy market law (REMIT): Regulation (EU) No 1227/2011, as amended by Regulation (EU) 2024/1106 (REMIT II), prohibits insider trading and market manipulation in wholesale energy markets. REMIT II also adds obligations for algorithmic trading. Forecasting systems must not use undisclosed inside information, and automated trading on forecasts falls under these rules [14][15].
• Financial regulation: MiFID II algorithmic-trading rules and MAR apply when forecasts drive automated trading in financial instruments (commodity derivatives). DORA applies to financial entities' ICT and third-party risk. Hedge accounting (IFRS 9) needs documented hedging strategies, and the forecast model becomes part of that documentation.
• Governance: a written procurement/hedging policy with limits, a model risk management process (validation, backtests, change logs), and segregation between forecasting and execution. Check data licences: many exchange and PRA licences forbid redistributing derived data.
• Typical risks: overconfidence in point forecasts, regime breaks, look-ahead bias, vendor black boxes, and herd behaviour. Mitigations: probabilistic outputs, benchmark against the futures curve, stress scenarios, human approval.

Getting started

1. Use spend analytics (DATA-015) to find the 3–10 commodities and energy carriers that drive most of your cost and margin risk, and write down how you buy today (tranches, fixing, index clauses).
2. Collect point-in-time history: your purchase prices and dates, the relevant indices/futures (e.g. SMARD and ENTSO-E for power, LME/PRA for metals) and key drivers.
3. Build a transparent baseline (futures curve, seasonal naive, LEAR/GBM) and test a zero-shot foundation model (Chronos-2 or TimesFM 2.5) in a backtest, or trial 1–2 Path A vendors on the same period.
4. Turn forecasts into decision rules (e.g. accelerate a tranche purchase if P50 rises above budget by more than X), approved in a procurement/hedging policy.
5. Run 3–6 months in shadow mode, comparing realised purchase prices with the "buy evenly" baseline, before you let the forecasts drive purchase timing.

Sources

1. Amazon Science — Chronos forecasting (Chronos-2, Chronos-Bolt; model sizes, licence, release dates) — https://github.com/amazon-science/chronos-forecasting
2. Google Research — TimesFM (versions 2.5 and 3.0, covariates, licences, BigQuery ML) — https://github.com/google-research/timesfm
3. epftoolbox — open benchmark and toolbox for electricity price forecasting (LEAR, DNN; EPEX-BE/FR/DE, Nord Pool, PJM) — https://github.com/jeslago/epftoolbox
4. Lago, Marcjasz, De Schutter, Weron — Forecasting day-ahead electricity prices: A review of state-of-the-art algorithms, best practices and an open-access benchmark, Applied Energy 293 (2021) — https://arxiv.org/abs/2008.08004
5. NEMO Committee — SDAC 15-minute MTU go-live (1 Oct 2025) — https://www.nemo-committee.eu/sdac
6. Bundesnetzagentur — SMARD market data platform — https://www.smard.de/en
7. Salesforce AI Research — uni2ts / Moirai 2.0 — https://github.com/SalesforceAIResearch/uni2ts
8. Volue — energy market data & SpotEx price forecasts — https://www.volue.com/
9. ChAI — AI commodity price forecasting and price-risk products — https://www.chaipredict.com/
10. Nixtla — TimeGPT documentation — https://www.nixtla.io/docs
11. EEX — market data and DataSource — https://www.eex.com/en/market-data
12. Nixtla — neuralforecast (models, exogenous support, Apache-2.0) — https://github.com/Nixtla/neuralforecast
13. Regulation (EU) 2024/1689 (AI Act) — https://eur-lex.europa.eu/eli/reg/2024/1689/oj
14. Regulation (EU) No 1227/2011 (REMIT) — https://eur-lex.europa.eu/eli/reg/2011/1227/oj
15. Regulation (EU) 2024/1106 (REMIT II amendment) — https://eur-lex.europa.eu/eli/reg/2024/1106/oj