chart-lookalike-transform 0.1.0

Find the affine transform that maps a candidate ticker's price series onto a familiar chart shape, and screen tickers that 'look like' a setup.

chart-lookalike-transform

Find the affine transform — vertical scale a, offset b, and optional inversion — that best maps a candidate ticker's price series onto a chart shape you already recognize, then screen a list of tickers by how well they "look like" that setup.

A chart's shape is invariant to vertical scale and shift: $10 → $11 over a month traces the same curve as $100 → $110. So to ask "does NVDA right now look like the cup-and-handle I'm watching for?" you fit

reference ≈ a · candidate + b          (a < 0 ⇒ candidate is the mirror image)

by least squares (closed form), report the fit quality as a 0..1 similarity, and overlay the transformed candidate on the reference with moving-average indicators.

This is not dynamic time warping — the time axis is held fixed, so "lookalike" means the literal same shape over the same window, just rescaled vertically. (DTW would match shapes that drift in time; deliberately out of scope.)

Install

pip install -e ".[test]"      # includes matplotlib for plotting + pytest
# or, runtime only:
pip install -e .              # numpy only; add ".[plot]" for PNG overlays

Python ≥ 3.10. Core dependency is just NumPy; matplotlib is optional (only for --plot).

Quick start (no network)

Everything works offline against deterministic synthetic data via --offline (or CHART_LOOKALIKE_API=mock), so you can try it instantly:

# List the built-in reference shapes.
chart-lookalike shapes

# Fit one ticker to a recognized shape and print the transform.
chart-lookalike match AAPL --shape cup-with-handle --offline

# Screen a basket, keep the best 5, and use a real ticker as the reference shape.
chart-lookalike screen AAPL MSFT NVDA TSLA AMD --ref-ticker NVDA --offline --top 5

# Write a PNG overlay (transformed candidate + reference + SMAs).
chart-lookalike match TSLA --shape v-bottom --offline --plot match.png

Drop --offline to pull live daily closes from Yahoo Finance (no API key needed). The data layer is provider-pluggable behind the CHART_LOOKALIKE_API env var.

Reference shapes

Three ways to specify the target shape (choose exactly one):

Flag	Meaning
--shape <name>	a built-in shape (chart-lookalike shapes lists them)
--ref-csv <path>	a file of reference prices (one number per line, or date,close rows — the last numeric field per line is used)
--ref-ticker <sym>	use another ticker's own series as the target shape

Built-in shapes: v-bottom, inverse-v-top, cup, cup-with-handle, head-and-shoulders, uptrend, downtrend, double-bottom.

How the fit works

For a reference t and candidate c (resampled to a common length), the least-squares affine fit has a closed form:

a = cov(c, t) / var(c)
b = mean(t) − a · mean(c)

When inversion is allowed, the mirror candidate −c is also fit and the lower-residual solution wins (a negative a ⇒ inverted = True). Fit quality:

rmse       = sqrt( mean( (a·c + b − t)² ) )
similarity = clip( 1 − rmse / rms(t − mean(t)),  0, 1 )

similarity = 1.0 is a perfect shape match; 0.0 is no better than predicting a flat line at the reference's mean.

Library API

import numpy as np
from chart_lookalike import fit_affine, apply_affine, get_provider, screen
from chart_lookalike.shapes import get_shape

provider = get_provider("mock")              # or "yahoo"
series   = provider.history("AAPL", days=180)
ref      = get_shape("cup-with-handle")

fit = fit_affine(series.closes, ref)         # AffineFit(scale, offset, inverted, rmse, similarity, n)
overlay = apply_affine(series.closes, fit)   # candidate mapped onto the reference's units

ranked = screen(ref, ["AAPL", "MSFT", "NVDA"], provider, top=3)
for r in ranked:
    print(r.symbol, round(r.similarity, 3), "inverted" if r.fit.inverted else "")

Live-data dependency & testing

The Yahoo provider hits query1.finance.yahoo.com/v8/finance/chart at runtime — that endpoint is undocumented and rate-limited, so it is the one piece that can break outside our control. The JSON parser (parse_yahoo_chart) is split out as a pure function and unit-tested against a saved fixture (fixtures/yahoo_aapl.json, including a null-close hole and the error envelope) — the test suite never touches the network.

pytest -q

Disclaimer

This is a charting/screening aid, not investment advice. A high similarity score means two curves have a similar shape over a window — nothing about what happens next.

License

MIT — see LICENSE.