Neurogica/forecast-workflow-bench

Forecast Workflow Bench (FWBench)

Language Model track: 1,251 capacity-planning cases across electricity and cycle hire.

Code · Leaderboard

Release: primary-1251-v2. This release promotes all four cohorts to the main evaluation. The earlier primary-377-v1 tag remains a separate reproducible result. Evaluation of all twelve language-model configurations on the expanded set is complete. Scores on 377 cases must not be presented as scores on 1,251 cases.

What is evaluated?

FWBench evaluates language models using a fixed harness and fixed forecast tools. The model chooses forecasts, context and horizon under a service budget, then submits eight three-hour capacity decisions. Real observations are paired with simulated shortage/surplus costs, a capacity grid and ramp limits. The service budget is 113,726 credits; the episode allows 12 tool calls and 16,384 output tokens. The language-model API price is not included in forecast credits.

Cohort	Cases	Target dates in 2026	Source
Electricity, July	377	July 24–27	EIA
Electricity, winter	339	February 3, 11, 19, 27	EIA
Electricity, August	375	August 3, 11, 19, 27	EIA
Cycle hire	160	May 3, 11, 19, 27 (London local days)	TfL

There are 50 electricity authorities and 20 cycle-hire stations. Cases contain 336 visible hourly values and a 24-hour target at one-day or seven-day lead. Histories end three hours before the decision time; required forecast horizons are 50 or 194 hours. London local days were converted from BST to UTC.

Ranking

For each case, let J = loss + lambda * credits / B, with B = 113,726. The default G is J minus the minimum J of the affordable eleven-policy hindsight menu. Lower G is better. The oracle uses the same common budget for every agent; spending more strictly increases G at fixed loss when lambda > 0. The default lambda is 0.01; sensitivity values are 0, 0.001, 0.003, 0.01, 0.03, and 0.1. This is an explicit reporting convention chosen after exploratory results, not an externally calibrated monetary conversion or preregistration. The hindsight menu is not deployable and need not contain the agent's plan; negative gaps are valid. Oracle credits count execution of its selected plan, not the offline cost of searching the menu.

Within each cohort, dates are averaged within series × lead, then these groups are equally averaged. Electricity and cycle hire each receive 50% weight; each of the three electricity cohorts receives 1/6 of the total weight. Loss, credits, validity, per-cohort results and case-weighted sensitivity are reported alongside G. A model is eligible only with complete outcomes on all 1,251 cases; operationally interrupted cases are not silently scored as model failures or excluded from a ranking.

Complete evaluation results

All twelve model configurations completed the same 1,251 cases. The primary report includes loss, forecast credits, validity and G at lambda = 0.01. Astra achieved the lowest agent G, while fixed Chronos-2 had lower observed loss and lower credits than Astra. Only fixed policies lay on the observed loss-credit Pareto frontier; these are sample-level comparisons. The paired interval for Astra minus fixed routing included zero.

• Complete model results
• Cost-weight sensitivity
• Pareto comparisons and exact coefficient intervals
• Paired uncertainty and lead/domain breakdowns
• Cohort reference results
• Decision quality by model group and loss-credit comparison

These result artifacts supplement the release; episode inputs, targets, tariffs and ranking rules are unchanged. The earlier release tag remains frozen; use the current main commit for this complete result bundle.

Download and layout

Use a resolved Hub commit SHA or the release tag to identify an exact snapshot:

hf download Neurogica/forecast-workflow-bench --repo-type dataset \
  --revision primary-1251-v2 --local-dir ./fwbench-data
python3 ./fwbench-data/examples/verify.py ./fwbench-data

Path	Contents
episodes/<case_id>/	Task, catalog, contract, tariffs, fixed model pool, history
reproduction_targets/<case_id>.json	Actual values and target timestamps for the offline scorer
manifests/cases.json	Complete case index with paths and component licenses
manifests/settings.json	Limits, aggregation and ranking definition
manifests/files.json	Payload sizes and SHA-256 checksums
viewer/cases.jsonl	Browsable metadata and instructions
provenance/	Source acquisition, terms, selection and previous release evidence
examples/	Integrity verification and a free classical-policy example
results/	Complete model metrics, sensitivity and uncertainty analysis

The viewer is an index; download the snapshot for executable inputs. Use the unit-aware FWBench SDK. For scoring, pass actual from the target JSON object. Only the current episode may be exposed to an agent. Do not mount the root, sibling histories, viewer or targets into its environment: overlapping windows can reveal another case's future. Folder separation alone does not enforce this.

Construction, licenses and limitations

Electricity observations are EIA-reported demand, not adjusted/imputed columns. The July cohort retains the original 377 cases. The winter and August cohorts apply the same completeness and quality rules to four separated dates each. The additional electricity instructions explicitly describe the existing free classical tool pool; this clarification is included in the distributed task bytes.

Cycle-hire data aggregate 1,665,846 completed TfL journeys from four April–May 2026 archives into hourly station departure counts. The 20 stations were selected by April 1–7 departure counts before forecasting results were inspected. Zeros were retained; no trip IDs, bicycle IDs or individual trip records are distributed. The target is recorded departures, not latent unmet rental demand. Capacity is a simulated service rate in departures/hour, not bike stock or a rebalancing plan. The restrictive shared action grid can yield positive loss even with perfect forecasts; report this floor separately.

Original benchmark contributions use Apache-2.0. EIA observations retain their source status. TfL-derived data retain the Transport Data Service terms, including attribution; they are not relicensed as Apache-2.0, CC0 or unmodified OGL. See LICENSE and the archived terms under provenance/extensions/tfl/.

Powered by TfL Open Data. Contains OS data © Crown copyright and database rights 2016. Geomni UK Map data © and database rights [2019]. No endorsement or rights to TfL branding, logos or personal data are implied.

Both sources are retrospective archives, not historical publication vintages. Observation dates after pinned forecast-model revisions reduce some training overlap risks but do not prove exclusion from every model's pretraining, including the language models. Public targets support reproduction, not unseen-test claims. Disclose any tuning or selection using these data. Case windows overlap; there are 639 series–target-day units, not 1,251 independent observations. This remains two domains with simulated contracts, not evidence for all operational settings.

Citation

Shunya Nagashima, Neurogica Inc. Forecast Workflow Bench: Evaluating Language-Model Decisions with Budgeted Forecast Tools. Manuscript in preparation. Please include the dataset revision and protocol version. No conference acceptance is claimed.