Inference LensWednesday, July 29, 2026

Machine intelligence for the crude complex.

CutPoint reads the primary sources of the energy complex — EIA, FRED, EDGAR, TSA — then trains forecasting models on them with the discipline of a backtest you can audit. You see the data, the folds, and the misses. No black boxes, no cherry-picked track record.

Open the Lens →read the methodology

The tape · FRED DCOILWTICO

WTI spot, the last 90 sessions.

Fig. 1 · Daily close · Source: FRED, U.S. EIALast 84.38 USD/bbl · 2026-07-20

The discipline

Every forecast is scored against what actually happened. In public.

Walk-forward backtests · point-in-time features·Sources: EIA, FRED, EDGAR, TSA

Most published “track records” are curve-fit hindsight. CutPoint's models are trained only on data that was publicly knowable at each moment — the join key is feature_available_at, not the observation date — then walked forward through history fold by fold. The model card shows every fold, every prediction, and every miss, including the days the 80% band failed to contain reality.

It's the standard a desk quant would hold an internal model to — published, with the receipts.

inspect the backtests in the lab →

The stack

Ingest. Train. Publish the record.

01Primary sources

Ingest

WTI · forecast vs actual · 30d

actualforecast

Scheduled pipelines pull EIA prices, FRED macro series, SEC EDGAR capex filings, and TSA throughput — each row stamped with when it became publicly knowable. Idempotent, windowed, logged.

EIAFREDEDGARTSA

02Quantile forecasting

Train

LightGBM quantile models fit p10/p50/p90 on point-in-time feature frames, validated with walk-forward cross-validation against a no-change baseline the model has to beat.

LightGBMQuantileWalk-forward

03Open model cards

Publish

Every trained model gets a public card: features, hyperparameters, all folds, all predictions, calibration — and a glossary that explains each metric in plain language.

Model cardsCalibrationGlossary

The lab

The models are public. So are the misses.

The lab lists every model CutPoint has trained — the config, the walk-forward folds, the days the band broke. Reader accounts browse the list; Library members open every card.

Visit the lab how the backtests work

Trained models · live from the registry

crack_321_change_7d_lgbm_v17838833307dMAPE 166.4%
wti_return_7d_chronos_bolt_v17838823657dMAPE 138.9%
wpsr_dist_direction_v17838822487d—
wti_return_7d_lgbm_v17838821507dMAPE 131.4%

Sources

Primary documents.
Nothing secondhand.

Every series on CutPoint traces to a government or regulatory primary source, ingested on a schedule and stamped with the moment it became public knowledge.

EIAU.S. Energy Information Administration — Petroleum prices & physical dataWeekly, Wednesdays

FREDFederal Reserve Economic Data — WTI & Brent spot, fed funds rateDaily, weekdays

EDGARSEC filings — E&P capital-expenditure disclosuresWeekly, Mondays

TSATransportation Security Administration — Passenger throughput (jet-fuel demand proxy)Daily