ggstyle

Publication finishing and safe date axes for Matplotlib.

v0.4 adds a production publication-finishing kit to the safe date axis: transactional labels, pure numeric labellers, immutable qualitative/continuous palettes, parameterized themes, collision-aware direct labels, inspectable dry runs, and deterministic figure export. Native Matplotlib plotting remains the intended path; there is no line() helper or general grammar compiler.

The date-axis behavior is tested, but the project is still young and follows semantic versioning. See the known limits before using collapsed mode in production.

Why

Most of the pain in Python time-series plotting is not the grammar, it's the axis: ticks in the wrong places, labels rotated to hide the fact that there are too many of them, weekend gaps shredding an intraday chart, and annotation code that quietly puts your vertical line three days off. ggstyle fixes the axis first.

Install

pip install ggstyle

For development from a clone:

pip install -e ".[dev]"

Use

It adopts any Axes, including plots it never made:

import matplotlib.pyplot as plt
import ggstyle as gs

gs.use_theme()                            # "minimal" is the default

fig, ax = plt.subplots()
ax.plot(df["date"], df["close"])          # plain matplotlib, seaborn, or df.plot()

gs.dates(ax).ticks("quarterly").fmt("month-year").zoom("2020", "2022")

Configuration and drawing methods return the handle, so calls chain.

Axis semantics are also available as structured data rather than only rendered output:

summary = gs.dates(ax).summary()
caption = gs.dates(ax).caption(add=True)

Finish labels

Apply a coherent title hierarchy and axis labels to the existing axes:

result = gs.finish(
    ax,
    title="Revenue",
    subtitle="Trailing twelve months",
    caption="Source: annual report",
    theme=gs.theme_spec("minimal", base_size=11),
    x=gs.axis(title="Date"),
    y=gs.axis(title="USD", labels=gs.label_currency("$", decimals=0)),
)

result.axes is exactly ax; all returned artists are native Matplotlib objects. Subtitle and caption artists participate in figure layout, repeated calls replace them in place, and False removes them. Use dry_run=True to validate and inspect the operation without mutation. finish() never edits data artists or global rcParams. When a theme is supplied, its diagnostic reports settings such as colour cycles and figure size that can only be applied safely before artists or figures are created.

Inspect the complete validated request without drawing or mutation:

plan = gs.finish(ax, title="Revenue", theme="minimal", dry_run=True)
payload = plan.as_dict()  # strict JSON-compatible plain values
print(plan.describe())    # stable formatted JSON

Replace a multi-series line legend with labels at the final visible data points:

ax.plot(x, revenue, label="Revenue")
ax.plot(x, forecast, label="Forecast")

result = gs.finish(
    ax,
    direct_labels=gs.end_labels(collision="avoid", fallback="legend"),
)

Endpoint labels use each line's colour, reserve figure space on the right, and separate nearby labels vertically without moving the data anchors. The operation is all-or-nothing: if a public legend entry is not a visible ordinary Line2D, its endpoint is outside the view, or the labels do not fit, the default policy builds a conventional legend instead. Use fallback="raise" to reject that plot during preflight, collision="none" to retain exact endpoint positions, and direct_labels=False to remove labels managed by ggstyle.

Save figures

Export an explicit figure with publication-oriented defaults and overwrite protection:

path = gs.save(
    fig,
    "report.png",
    width=7,
    height=4,
    units="in",
    dpi=300,
    metadata={"Creator": "ggstyle"},
)

Both dimensions are required, units may be in, cm, mm, or px, and the format is inferred from the suffix unless supplied explicitly. Output is opaque and tightly bounded by default; use transparent=True or bbox="standard" explicitly when needed. Tight bounds crop the requested canvas to its decorated content, while standard bounds retain the exact canvas dimensions.

save() refuses to overwrite by default. With overwrite=True, it renders to a temporary file and replaces the destination only after success. A renderer failure therefore leaves an existing file intact, and the figure's original size is restored in every case. SVG IDs are stable and variable SVG/PDF timestamps are suppressed by default.

Ticks — where they go

.ticks("monthly")                   # daily | weekly | monthly | quarterly | yearly
.ticks("month-end")                 # anchored: month-start, quarter-end, year-start, ...
.ticks(every="3M")                  # any offset alias; legacy M/Q/Y/H accepted
.ticks(n=6)                         # about six ticks, snapped to a natural cadence
.ticks(at=["2020-01-01", "2021-07-01"])
.ticks(major="yearly", minor="monthly")

Anchoring is not cosmetic: month-start vs. month-end is the difference between labels that line up with your observations and labels that float between them.

Labels — what they say

.fmt("concise")      # default: year shown once, not on every label
.fmt("month-year")   # Jun 2020
.fmt("quarter")      # Q2 2020
.fmt("year") / .fmt("month") / .fmt("day") / .fmt("iso") / .fmt("time")
.fmt("%b '%y")       # any strftime string
.fmt(lambda d: f"week {d.isocalendar().week}")

Changing the format never moves a tick, and changing the cadence never changes the format. That orthogonality is a test, not an aspiration.

Numeric axes

Percent, currency, grouped-number, and SI-prefix labels are locale-independent callables:

currency = gs.label_currency("$", scale=1_000_000, decimals=1, suffix="M")
ax.yaxis.set_major_formatter(gs.as_formatter(currency))

gs.label_percent(decimals=1)(0.125)   # "12.5%"
gs.label_number(decimals=2)(1234.5)  # "1,234.50"
gs.label_si(unit="B")(1_500_000)     # "1.5 MB"

Factories do not mutate Matplotlib. The explicit adapter returns an ordinary matplotlib.ticker.FuncFormatter, so axes and formatter objects remain directly available.

Palettes

The palette API exposes the shared eight-colour theme cycle and perceptually ordered continuous options without changing Matplotlib configuration:

ax.set_prop_cycle(color=gs.palette("qualitative").colors)

colors = gs.palette("sequential", n=5).colors
neutral = gs.palette("diverging").at(0.5)  # "#F7F7F7"

Qualitative requests above eight fail instead of manufacturing ambiguous colours. Diverging samples require an odd count so their explicit neutral midpoint is retained. Continuous lookup makes missing and out-of-bounds behavior explicit through missing_color= and out_of_bounds=. Palette values are immutable and can be passed to ordinary Matplotlib cycles and colormaps.

Range

Partial strings expand to whole periods, pandas-style:

.zoom("2020", "2022")      # three complete years
.zoom("2020-03", None)     # open-ended
.zoom(last="6M")           # trailing window from the last observation, not from today
.zoom(ytd=True)
.pad(left="1M", right="1M")

Gaps

.collapse()   # unobserved dates get no space
.expand()     # true datetime axis, gaps restored

Collapsed mode is defined by the dates present in your data, not by a holiday calendar. Anything not observed is not allocated space. That is correct for any market or region and needs no extra dependency. With several series, the axis uses the union of observed dates.

Annotation in date space

Every one of these is correct in both modes — that is the whole point of the handle:

.loc("2020-03-23")                    # -> native matplotlib date coordinate
.vline("2020-03-23", label="trough")
.span("2020-02-19", "2020-03-23", label="drawdown")
.spans(events_df, start="begin", end="end", label="name")
.clear_annotations()                    # remove managed annotation artists safely
.grid("yearly")                       # gridline cadence, independent of ticks

In collapsed mode the scale places a date inside a gap (a Sunday, a holiday) by linear interpolation between its neighbours. loc() always returns the same native matplotlib date coordinate in either mode; loc(date, snap=True) rounds to the nearest observation, and loc(date, strict=True) raises if the date was never observed.

Escape hatch

Native matplotlib date input now passes through the same registered scale. Use datetime values directly, or use .loc() when you want ggstyle's parsing, snapping, or strict lookup:

handle = gs.dates(ax).collapse()
ax.axvline(pd.Timestamp("2020-03-23")) # lands in the right place
ax.set_xlim(handle.loc("2020-01"), handle.loc("2021-01"))

Themes

Nine ggplot2-inspired themes ship. minimal remains the default.

gs.use_theme()             # minimal, process-wide
gs.use_theme("grey")       # "gray" also accepted
gs.use_theme("bw")

with gs.theme("dark"):     # scoped; restores every rcParam on exit
    ...

plt.style.use(gs.stylesheet())   # the .mplstyle on its own, no ggstyle import needed

Create a reusable recipe when a report needs a different type scale, family, or a small set of Matplotlib overrides:

report_theme = gs.theme_spec(
    "minimal",
    base_size=11,
    base_family="DejaVu Sans",
    overrides={"axes.titlesize": 14},
)

with gs.theme(report_theme):
    fig, ax = plt.subplots()      # complete creation-time styling

gs.finish(ax, theme=report_theme) # safe non-data styling on an existing axes
params = gs.theme_params(report_theme)  # pure, read-only resolved mapping

Recipes validate names and values immediately. Base sizing scales the full theme type system proportionally, base family is applied next, and explicit overrides win. Applying a recipe through finish() updates panel and figure surfaces, spines, grid lines, ticks, titles, labels, and an existing legend. It deliberately preserves data artists, property cycles, figure geometry, line defaults, save settings, and global rcParams; those creation-, data-, and output-time settings are listed in result.diagnostics.

Available names are minimal, grey, bw, linedraw, light, dark, classic, void, and test. The corresponding ggplot2 function spellings, such as theme_bw and theme_classic, are accepted as aliases. test is intended for stable visual tests, while void removes the plotting surface for maps and other annotation-free displays.

All themes spell out the same type scale and public qualitative colour cycle, so switching changes the non-data surface rather than the plot's identity. The cycle is Okabe–Ito-derived, uses black in place of grey for stronger separation from the dark-theme surface, and is capped at eight; past eight, direct labelling or faceting is the right answer, not a ninth colour.

Importing ggstyle never mutates rcParams. Theming is always something you ask for.

Almost all of it is plain rcParams in a .mplstyle file, including spine removal (axes.spines.left: False), which an earlier draft of the design wrongly assumed needed Python. Facet-strip styling has no core matplotlib equivalent, and transparent axis labels in void may still reserve layout space.

Data frames

pandas and polars both work, as do pyarrow arrays, numpy datetime64, and plain lists:

gs.dates(ax, data=frame["date"])     # pandas Series, polars Series, or Index

Polars is detected by module name rather than imported, so installing ggstyle never pulls it in and pandas-only users pay nothing for the support. Timezone-aware input from either library is converted to UTC instants for positioning; display timezones stay a separate concern handled by .tz().

Two things are deliberately not guessed: a whole DataFrame passed where a column was meant, and a string column that might be dates. Both raise.

Missing values in explicit date data also raise unless exclusion is requested with missing="drop". The number excluded remains available through .summary() and in generated captions.

Multiple panels

Synchronize comparable axes with a live observation registry and common limits:

handles = gs.sync_dates(axes, mode="collapse", limits="union")

This prevents the same date from receiving different ordinal positions in independently collapsed panels. The handles share one revisioned registry: calling .refresh() on any member rescans every live member and updates every collapsed scale transactionally.

Call .refresh() after adding, changing, or removing plotted artists. A repeated gs.dates(ax) call also refreshes an existing handle. Call .dispose() to disconnect a handle and release its registry and managed-artist references without removing artists from the Matplotlib axes.

Design rules

• The date axis is a standalone object, not a side effect of plotting.
• Importing the package is inert; theming is opt-in.
• Placement, labels, gridline cadence, and range are four independent knobs.
• Fail loudly: a non-date axis raises, and mixed tz-aware/naive input raises rather than guessing UTC.
• Never resample or interpolate the data silently.
• Never rotate tick labels by default. Rotation is a symptom of bad tick selection.

Known limits

• Collapsed mode supports lines, scatter, fill_between, and native data-space or x-data blended transforms without rewriting their geometry. Lines, scatter collections, and native fill_between polygons contribute observations automatically. Because Matplotlib does not retain the source x array for step="mid", that form still requires the complete dates through gs.dates(ax, data=dates).
• Data artists with custom x transforms are rejected during refresh; use ax.transData because explicit dates cannot make a non-data transform safe.
• Unsupported or ambiguous date-bearing artists raise DateDiscoveryError during preflight. Version 0.4 retains the strict policy and has no permissive warning mode.
• .tz() assumes naive data is UTC when converting for display.
• Palettes map normalized values and select colours; data-domain training, category assignment, legends, and colorbars remain ordinary Matplotlib work until semantic scales land.
• The current finish() surface coordinates plot, subtitle, caption, axis-title, numeric-label formatting, safe existing-axes theming, and direct labels for ordinary Cartesian Line2D series. General label repulsion, scatter endpoint labels, and guide layout remain later work; filesystem export is intentionally separate in save().

Tests

python -m pytest -q
ruff check .
mypy

See CONTRIBUTING.md for the complete development workflow and SECURITY.md for vulnerability reporting.

The executable publication and nine-theme figures are in the documentation gallery; regenerate their reviewed assets with python tools/validate_gallery.py --write. The complete HTML documentation is published through GitHub Pages.

The structured documentation follows the same user-guide, API-reference, pitfalls, and release-note separation used by statsmodels. Build it locally with:

pip install -e ".[docs]"
python -m sphinx -W --keep-going -b html docs/source docs/build/html