The Power Buyer's Club

Setup Forecasts the next 4 days of electricity demand from 3 weeks of history.

Pros Runs offline in your browser FAST · ~5 MB one-time download

Usage Play the power stock market - buy power before the price moves!

So we got a stock market here?

Imagine you are a power buyer at a utility. Your job is to buy enough electricity to serve your customers, at the best price you can. Those Data Centers sure do make life hard! This model reads a 512 hour window of a real meter's past demand (about 3 weeks) and predicts the next 96 hours (4 days) in one shot. That forecast tells you when demand is about to climb, so you can buy before the market catches up. The trading terminal on the right simulates how a buyer would act on it.

Power Buying with HFT

The operator lives in the rFFT spectrum. You've heard this spiel before: a handful of spectral convolutions over the demand window replace a heavy duty sequence model. It streams with our handy dandy constant memory ring buffer, so a 512 sample window is the entire retained history (about 2 KB) and cost stays flat no matter how long the meter has been running. A forecast costs about 34 MFLOP of compute (an Atari could do it) and completes in a few milliseconds on one CPU core (measured live here!). At that speed a buyer can reforecast on every price tick; a forecast that takes seconds leaves the desk trading on a stale view - in stock terms, that's alpha.

Full card on Hugging Face ↗

Interactive demo

You are a utility power buyer. The model forecasts demand in a few ms. See the ramp before the market does, and buy power at the low price instead of into the spike.

Spot price

$0.0

illustrative

Your position

0 / 40 MWh

nothing bought yet

Your forecast

0 kW

real model

Slow desk sees

0 kW

same model, stale

Signal

...

waiting

prices below are illustrative · the forecast is real

forecasting demand…

demand forecast (real, next 96 h) demand that happened spot price (illustrative) your buys

Pick a load zone above to start a buying session.

Rounds played

buying sessions

Slow desk paid

forecast too slow

You saved

acting on a fast forecast

Live compute