Agentic RT−DAM spread forecaster · HB_NORTH hub ·
priors fed to the LLM: 14-day regime mean + LR direction model + TSBandit72 win-rate + spike-risk LR ·
no post-process harness — the LLM reasons over the priors and decides each hour itself ·
fixed 1 MW · ±$2/MWh threshold · runs daily 3 PM CT after DAM clears
How did we trade? —
Realized RT-DAM spread ($/MWh, right axis)
Tomorrow's positions —
Long DART (buy DAM, sell RT)
Short DART (sell DAM, buy RT)
Forecast spread ($/MWh, right axis)
Per-hour signal breakdown
| HE |
DAM $ |
Quant priors |
LLM decision |
Position MW |
Why — LLM rationale |
Persistence |
| Bandit |
LR |
Spread |
Conf. |
Spread |
MW |
Agent reasoning
—
Cumulative P&L — agent vs persistence