BatteryArb is built around public ISO data, battery-relevant derived metrics, and operating workflows that care about interval timing, outage posture, congestion pressure, and simulation before deployment.
BatteryArb uses public ISO feeds from ERCOT, PJM, MISO, ISO-NE, and IESO, with live-first support for ERCOT and PJM and normalized fallback paths for the rest.
Forecast responses combine intraday profile, load drift, outage trajectory, and weather stress. Forecast endpoints expose `model_version` and `mae_proxy_usd_mwh` for interpretability.
The goal is not generic data warehousing. The goal is to help battery teams make better decisions under short-horizon market uncertainty.
Typical development benchmark: 7-11 USD/MWh MAE on stressed and non-stressed 5-minute windows combined.
Typical development benchmark: 8-13 USD/MWh MAE depending on node granularity and outage regime.
Forecast band width and outage risk score should shape aggressiveness, not just point forecast direction.
Battery strategy evaluation compares baseline threshold dispatch to outage-aware bid/offer thresholds. We track MAE proxy, gross margin delta, congestion avoidance, and imbalance exposure reduction.
The simulator is built to improve as the market path evolves. It recalibrates charge and discharge posture every five minutes from rolling trend and realized volatility rather than assuming one static bid/offer pair survives the whole operating day.
The goal is not to claim perfect foresight. The goal is to package the context a strong operator monitors manually so the broader desk can make better interval decisions with less inference work.
Commercial pilots are usually judged on baseline-relative outcomes such as 2-8% dispatch margin lift, 5-15% fewer avoidable bad intervals, or faster decision cycles. Those are evaluation targets, not guaranteed returns.
Production outcomes can be posted back through the API. BatteryArb groups those observations by ISO, node class, time bucket, outage regime, and operator posture, then tracks forecast error, realized PnL, lift versus baseline, and win rate for each posture.
Learned adjustments remain tenant-scoped, require minimum sample counts before activation, and stay bounded within a narrow bid/offer adjustment range. That keeps the adaptive layer explainable and avoids overfitting to a few strong or noisy intervals.
Store keys in a secret manager or environment variables. Keep raw keys out of browser bundles. Use a server-side proxy for dashboard products.
BatteryArb supports issuance, revocation, email-only key delivery, and rotated resend flows so support workflows do not require raw key retrieval.