Models that anticipate prices, demand and risk — and say how confident they are. Uncertainty gets measured; it doesn’t get makeup.
Saying "the price will be 320" without saying how far it can deviate helps nobody who has to sign. A calibrated interval does: it bounds the risk you’re taking with the decision.
Nearly all of a year’s cost concentrates in a few extreme hours. A model that nails the average and misses the spikes is a model that failed you on the day it mattered.
There are decades of public data available and unexploited. Fetching it, cleaning it and turning it into a usable signal is half the job.
Straight from the official API, not from a file someone exports by hand. If the data isn’t reproducible, neither is the model.
Seasonal decomposition and clustering before modeling: understanding what moves the series avoids models that memorize instead of learn.
Conformal prediction to produce intervals with real coverage, not decorative error bars.
Spikes are modeled separately, with Markov chains for the regimes. It’s where an average model always fails.
What we use
Real projects from this service.
Spot-price forecasting with spike classification and calibrated intervals.
Decomposition and clustering of 26 years of system demand.
MILP optimization deciding, hour by hour, between selling or burning the fuel.
It depends on the phenomenon’s seasonality, but as a rule you need several full cycles: for something with yearly seasonality, three or four years. You can start with less, though the model will say little about what it hasn’t seen yet.
That when the model says "80% confidence", the real value falls inside 80% of the time. It sounds obvious and almost no model delivers it: it takes conformal prediction, validating coverage against data the model never saw.
Yes. The method — fetch the data from its source, separate signal from noise, quantify the uncertainty — is the same for product demand, portfolio risk or inventory turnover. What changes are the variables, not the approach.
[ Got something that should be running on its own? ]
Reply within 24 hours · Colombia