256M-parameter time series model t0-beta hits top-3 on GiftEval and FevBench, +$15M in ERCOT backtest
fpedregosa · x · 2026-09-17
ForecastingCO released t0-beta, its best model yet, with a blog post and tech report.
Key facts:
- Only 256M parameters, top-3 on both major time series forecasting benchmarks, GiftEval and FevBench
- Natively handles historical and future known covariates; improves on June's t0-alpha across univariate data and covariate handling
- Independent energy-sector benchmark: a 2-year backtest trading the ERCOT day-ahead market at 100MW per hour placed its models 2nd and 3rd with +$15M P&L
The model is available to try on your own data.
More from Venture
- AI-powered language learning platform lists for $7.5M with $137k MRR and 70% margins — marclou · 2026-09-18
- Baylor CIO: capital velocity beats multiple — three 3X funds over 18 years beat one 15X — MartinGTobias · 2026-09-17
- Stripe lays out AI payments stack: Link agent wallet, MPP, tokens and usage billing — jeff_weinstein · 2026-09-17
- An agent workflow for Reddit traffic: review 700 threads, reply to only 15, human approval before posting — mhdfaran · 2026-09-17
- Nebius GPU rate hikes and Fed hikes are connected: AI debt is pushing up yields — kevinsxu · 2026-09-17
- CNBC: OpenAI and Anthropic generate roughly 10x the revenue of all Chinese AI models combined — Polymarket · 2026-09-17