Home

dtfit documentation

dtfit fits models that are nonlinear in their parameters -- exponentials, sinusoids, logistic curves, saturating and peak shapes -- to noisy data, and recovers the physical parameters (a growth rate, a frequency, an asymptote), not just an opaque curve. It does this through differential (non-Taylor) transformations: instead of comparing the model to the data sample-by-sample, it compares their images -- a fixed-size statistic of projections onto a basis. The image is additive over samples, so one fit runs in one shot, chunk by chunk, across workers, or one sample at a time.

This folder is the documentation. Pick the door that matches what you need:

If you want to...	Read
Understand the ideas from scratch, no heavy math assumed	guides/ -- plain-language explanations of every method, with the proofs built up gently
See the full map -- every method, version, variant and adaptation	guides/lineage-and-variants.md -- the complete atlas of where each approach came from and how it relates
How the method evolved -- the stages from spectrum balance to the image, each one runnable	Lineage -- where every historical stage lives in the dtfit-legacy package
Look up a function or class -- signatures, arguments, return types	api/ -- complete reference for the public dtfit API
See the rigorous math -- the formal derivations and proofs	methods/ -- the mathematical reference, one file per method
Learn by running code -- copy-paste examples	examples/ -- quickstart -> methods -> models -> sklearn -> streaming -> scaling -> diagnostics
Understand the research -- the experimental adaptations and how they were validated	experimental/ -- the dtfit-experimental package, the adaptations in trial, and every baseline the methods are compared against
See what is measured -- every claim with its numbers, wins and losses	Experiments -- the experiment notebooks with a summary of each, and Domains -- the same results by application

The shortest possible introduction

There are three core fitting methods, plus a streaming one. They are all the same idea (match the images of the model and the data) applied differently:

• LSI -- the batch fit in the Legendre basis: accurate and general-purpose, at parity with pointwise least squares. Start here.
• EAC -- the batch fit in the block basis of equal windows: local, the basis for records with jumps at known epochs and for data known only as per-window totals.
• DSB -- a symbolic reference method, used to derive and check the others; not for production.
• ImageFilter -- the streaming version: feed one sample at a time, track parameters that change over time, and detect when the system changes regime. ImageFilter(basis="legendre") and ImageFilter(basis="block") fix its basis to Legendre and block.

On top of those sit convenience layers: a scikit-learn estimator (NonlineRegressor), a model catalog so you pick a shape instead of writing a formula, one-call "just fit it" entry points (fit(..., basis="auto"), auto_forecast), and scaling backends for big or multi-channel data.

For genuinely random data (economic / financial series) there is a dedicated stochastic-series solution: it fits the deterministic functionals of the process (its autocovariance, spectrum, trend/cycle) to characterize, forecast, and even generate it (fit_stochastic, Stochastic), with a streaming twin (StochasticFilter) that tracks the structure live.

New here? Open guides/README.md.