What's New

Stay up to date with every PyTrendy release — user-facing improvements, bug fixes, and behaviour changes.

---

Agentic Docs, Noise Control & Trend Fixes (v1.2.0)

Released 2026-05-09

Four updates in v1.2.0: an agentic docs generator, a new noise toggle, and two fixes to trend metrics and normalised input handling.

Automated What's New — agentic docs generator

The What's New page itself is now generated by an AI agent. A GitHub Actions workflow (whats-new.yaml) fires automatically whenever a GitHub Release is published. It invokes scripts/generate_whats_new.py, which calls the GitHub Models API (Claude Sonnet 4.6) to write a user-friendly What's New entry from the release notes, then opens a pull request for review. Introduced: #120

Key behaviours:

• Pre-releases (from develop) add a collapsible entry to the Upcoming Changes section.
• Stable releases (from main) add a versioned entry and open a sync PR back to develop.
• Code examples reference GitHub raw URLs (develop for pre-release, main for stable) so they are always reproducible.
• Within each section, fixes and features are listed in descending time order (most recent at the top).
• Before/after plot comparisons use plot_pytrendy(df, value_col, segments, suppress_show=True) — this ensures a consistent figsize=(20, 5), weekly grid, and legend across all images so they are directly comparable.
• Agent instructions embedded in the script docstring remind the generator to verify that referenced CSV files still exist before each refresh.

Noise detection control (avoid_noise)

A new avoid_noise parameter in method_params lets users opt out of noise detection entirely. When set to False, spikes and noisy regions are ignored and trend detection proceeds straight through them. Introduced: #110

Useful for modelling a new-market launch or quasi-experiment where the signal is zero before and after the activation window. With avoid_noise=False, boundary artifacts around step-changes are suppressed, yielding clean Up/Down segments.

Before — avoid_noise=True (default)

After — avoid_noise=False

Code

import pytrendy as pt

df = pt.load_data("series_synthetic")
df.set_index("date", inplace=True)
# Simulate a new-market / quasi-experiment: zero activity before and after activation
df.loc["2025-01-01":"2025-02-27", "abrupt"] = 0
df.loc["2025-05-05":"2025-06-30", "abrupt"] = 0
df = df.reset_index()

result = pt.detect_trends(
    df, date_col="date", value_col="abrupt",
    method_params=dict(is_abrupt_padded=True, avoid_noise=False)
)
print(result.df[["direction", "start", "end"]])

Metrics for all segment types

Segment metrics (pct_change, change_rank) were not computed for every trend type. All output rows now carry complete metric columns regardless of classification. Fix: #88

direction	start	end	pct_change	change_rank
Up	2000-01-01	2000-05-01	+42%	1
Down	2000-06-16	2000-12-31	−38%	2
Flat	2000-05-02	2000-06-15	+0%	3

Before v1.1.11, all three columns were NaN for Flat (and Noise) segments. Ranked by change_rank.

Code

import pandas as pd
import pytrendy as pt

url = "https://raw.githubusercontent.com/RussellSB/pytrendy/main/tests/tests_crashes_edgecases/data/low_value_series.csv"
df = pd.read_csv(url)
result = pt.detect_trends(df, date_col="date", value_col="trend", plot=False)
print(result.df[["direction", "pct_change", "change_rank"]])

Trend detection on normalised time series

detect_trends() previously returned an empty result when the input signal was scaled to the [0, 1] range (e.g., after min-max normalisation). The absolute detection threshold was too large relative to the signal amplitude — causing the algorithm to detect nothing, which the flat fill-in logic then represented as a single all-Flat series. Fix: #79

Before — v1.1.10

After — v1.1.11

The same normalised series now returns a correctly detected Up / Flat / Down sequence. Regression test: test_low_value_series

Code

import pandas as pd
import pytrendy as pt

url = "https://raw.githubusercontent.com/RussellSB/pytrendy/main/tests/tests_crashes_edgecases/data/low_value_series.csv"
df = pd.read_csv(url)
result = pt.detect_trends(df, date_col="date", value_col="trend")
print(result.df[["start", "end", "direction"]])

---

Noise Detection & Robustness (v1.1.3 – v1.1.10)

A sustained series of improvements to noise detection, spike precision, and edge-case stability — making the algorithm significantly more reliable on real-world noisy signals.

Released in v1.1.10

Released 2026-03-21

Comprehensive automated tests added for noise edge cases and crash scenarios — full coverage for the noise detection module. (#46)

Details

• Automated tests for noise crashes (test_noise_crashes.py) and edge cases (test_noise_edgecases.py).
• Artifact-cleaning helpers refactored to be more testable and deterministic.
• Several pytest-mpl baseline images corrected.

---

Released in v1.1.8 and v1.1.9

v1.1.8 — 2025-11-15 · v1.1.9 — 2026-02-07

Targeted improvements to noise detection precision and flat segment handling. (v1.1.8: tag · v1.1.9: #42)

Flat fill-in improvements (v1.1.8)

Building on the flat fill-in first introduced in v1.1.0, v1.1.8 extended coverage to two additional edge cases:

• Trailing / leading regions: if the detected segments don't span the full series range, the uncovered leading or trailing period is now filled in with a Flat segment.
• Robustness to grouping: zero-day boundary regions are safely skipped; grouped segments no longer produce gaps.

The before/after below shows three spikes distributed across a gradual series. In v1.1.7, the period after the last Noise segment (2025-06-18 → 2025-06-30) was left uncoloured. In v1.1.8 it is correctly filled as Flat.

Before — v1.1.7

After — v1.1.8

Regression test: test_gradual_four_spikes_distributed_flatfillins

Code

import pandas as pd
import pytrendy as pt

df = pt.load_data("series_synthetic")
df["date"] = pd.to_datetime(df["date"])
df.set_index("date", inplace=True)
df.loc["2025-04-08":"2025-04-08", "gradual"] = 200
df.loc["2025-05-08":"2025-05-08", "gradual"] = 200
df.loc["2025-06-08":"2025-06-08", "gradual"] = 200
df = df.reset_index()
pt.detect_trends(df, date_col="date", value_col="gradual",
                 method_params=dict(is_abrupt_padded=True))

Noise detection (v1.1.8)

• Better precision for a spike on an otherwise flat-zero signal.
• Improved sensitivity when flat conversions emerge from noisy gradual trends.
• trend_too_flat now treated as a flat conversion rather than noise.
• Up/down classification uses the actual signal (not the smoothed derivative), reducing downstream artifact-cleaning needs.
• Gradual trends enclosed in noise are handled more leniently.
• Noise adjustment for contract logic now crops correctly before start/end boundary checks.

Spike precision (v1.1.9)

Further improvement to spike detection precision for signals with a single dominant outlier surrounded by otherwise stable values. (#42)

---

Released in v1.1.3 – v1.1.7

v1.1.3 — 2025-10-16 · v1.1.4 — 2025-10-19 · v1.1.5 — 2025-10-22 · v1.1.6 — 2025-10-23 · v1.1.7 — 2025-11-01

A focused series of noise detection improvements, from an initial major revamp through to edge-case tuning and stability fixes.

v1.1.7 — expand-contract & noise stability

(tag v1.1.7)

• Expand-contract: gradual trends can now be retroactively updated when a newer gradual changes the reference baseline. (a99c30f)
• Noise detection: resolved edge cases around abrupt-noise boundaries, opposite-direction overlaps, and post-grouping validity checks. (9b41189, 06aa45c)
• Plot: visual displacement is only applied when it does not break the up/down direction contract. (7ae27ad)

v1.1.6 — noise threshold tuning

Made the noise threshold slightly less sensitive to avoid false positives on near-zero signals. (#16)

Signals where all values are very close to zero were being incorrectly classified as alternating Noise/Flat bands. After the threshold adjustment, the same signal is correctly identified as a single Flat segment.

Before — v1.1.5

After — v1.1.6

v1.1.5 — abrupt shaving infinite loop

Fixed an infinite loop in abrupt shaving when a segment was broken into abrupt sub-segments. (#14)

v1.1.4 — noise detection major revamp

Trend detection now much less sensitive to noise spikes overall. Introduces DTW-based abrupt/noise distinction and more robust spike classification. (#13)

Before the revamp, the algorithm fragmented noisy-but-flat segments into many small alternating Noise/Flat bands and missed underlying gradual downtrends. After, it consolidates the noise and correctly identifies the downtrend structure.

Before — v1.1.3

After — v1.1.4

Regression test: test_noisy_edgecase_3_scenario

Code

import pandas as pd
import pytrendy as pt

url = "https://raw.githubusercontent.com/RussellSB/pytrendy/main/tests/tests_crashes_edgecases/data/noisy_edgecases.csv"
df = pd.read_csv(url)
pt.detect_trends(df, date_col="date", value_col="noisy_edgecase_3")

v1.1.3 — spikes on gradual trends

Improved handling of spike segments that sit on top of gradual trends. (#12)

Previously, a single isolated spike in a gradual trend series was expanded into a wide Noise band spanning several weeks, masking the surrounding Up/Down structure. After the fix, spikes are classified as a tight 1–3 day Noise segment with the gradual trends fully intact on either side.

Before — v1.1.2

After — v1.1.3

Regression test: test_gradual_spike_single_mid_series

Code

import pytrendy as pt

df = pt.load_data("series_synthetic")
df.set_index("date", inplace=True)
df.loc["2025-03-25":"2025-03-25", "gradual"] = 200  # inject spike
df = df.reset_index()
pt.detect_trends(df, date_col="date", value_col="gradual",
                 method_params=dict(is_abrupt_padded=True))

---

Core Engine & Initial Launch (v1.0.x – v1.1.2)

The foundation of PyTrendy — from initial release through the first major engine overhaul that introduced flat fill-in, a cleaner results API, and comprehensive robustness improvements.

Released in v1.1.1 and v1.1.2

v1.1.1 — 2025-10-15 · v1.1.2 — 2025-10-15

Patch releases addressing deployment pipeline issues and a relative-import fix introduced when v1.1.0 restructured the package layout. No user-facing behaviour changes. (v1.1.1 · v1.1.2)

---

Released in v1.1.0

Released 2025-10-15 — minor version: new features and major robustness overhaul

The most significant update to PyTrendy's core engine since the initial launch — new flat fill-in, a simpler results interface, and a thorough revamp of the signal processing pipeline. (#8)

Flat fill-in

Gaps between classified segments are now automatically filled with a Flat segment, so the output always covers the full input time range.

Before this change, those uncovered intervals appeared as white gaps between coloured regions. After v1.1.0, they are explicitly represented as blue Flat segments.

Before — v1.0.x

After — v1.1.0

Regression test: test_gradual_four_spikes_distributed_flatfillins

Code

import pytrendy as pt

df = pt.load_data("series_synthetic")
df.set_index("date", inplace=True)
df.loc["2025-02-28":"2025-02-28", "gradual"] = 125
df.loc["2025-04-09":"2025-04-09", "gradual"] = 150
df.loc["2025-05-08":"2025-05-08", "gradual"] = 300
df.loc["2025-06-03":"2025-06-03", "gradual"] = 320
df = df.reset_index()

pt.detect_trends(
    df,
    date_col="date",
    value_col="gradual",
    method_params=dict(is_abrupt_padded=False),
)

Abrupt trend detection — illustrated

v1.1.0 introduced is_abrupt_padded for the first time, enabling boundary padding so that abrupt transitions span their natural width rather than hairline boundaries. This unlocks quasi-experiment designs such as Interrupted Time Series Analysis (ITSA), where clean pre/post intervention windows are required.

Before — v1.0.x

After — v1.1.0

Code

import pytrendy as pt

df = pt.load_data("series_synthetic")
pt.detect_trends(df, date_col="date", value_col="abrupt",
                 method_params=dict(is_abrupt_padded=True))

Simplified results interface

The TrendyResults object gained two shorter access patterns:

Before (v1.0.x)After (v1.1.0+)

segments_df = result.segments_df
summary_df  = result.summary["df"]

segments_df = result.df          # shorter alias
summary_df  = result.df_summary  # direct attribute

Both names continue to work in v1.1.x.

Brown-bug fix

Up (green) and Down (red) regions were stacking on top of each other in v1.0.x, blending into a brownish artifact. The root cause — abrupt spikes being shaved without direction-awareness — was resolved; segments now align directionally before any visual displacement. (b0d1690 · 5509178)

Before — v1.0.x

After — v1.1.0

Code

import pytrendy as pt

df = pt.load_data("series_synthetic")
df.set_index("date", inplace=True)
df.loc["2025-01-01":"2025-02-11", "abrupt"] = 0
df.loc["2025-02-16":"2025-03-10", "abrupt"] = 125
df.loc["2025-03-18":"2025-04-15", "abrupt"] = 150
df.loc["2025-03-20":"2025-04-22", "abrupt"] = 250
df.loc["2025-03-25":"2025-04-01", "abrupt"] = 200
df = df.reset_index()

pt.detect_trends(
    df,
    date_col="date",
    value_col="abrupt",
    method_params=dict(is_abrupt_padded=False),
)

Other core processing revamp & bug fixes

The signal processing and post-processing pipeline was extensively reworked (#8 · 6c53790):

• More robust handling of edge cases across all trend types.
• Abrupt detection: better shaving, sub-segmentation, and direction-sensitivity (c66ed2e).
• Gradual swallowing stretches flexibly across neighbouring segment adjustments (5f6e2c9).
• Touching consecutive abrupt segments are now correctly grouped (50d2f52).
• has_inverse() now also validates total-change consistency, not just direction (37411fe).
• Windows path separators handled correctly in the data loader (d04c49c).
• detect_trends() is now robust to wide DataFrames containing non-numeric columns (a7c286d).
• Fixed a crash when no segments are detected (a3ac41f).

---

Released in v1.0.x

August–September 2025 — initial release

PyTrendy launched with gradual, abrupt, and flat trend detection in a single call. (v1.0.0)

Gradual trend detection

Smooth up and down trends are detected and annotated as Up, Down, and Flat segments out of the box with a single call to detect_trends().

Code

import pytrendy as pt

df = pt.load_data("series_synthetic")
pt.detect_trends(df, value_col="gradual", date_col="date")

Abrupt trend detection

Abrupt step-changes in the signal are detected and annotated as Up or Down regions. v1.0.x already supported multi-segment abrupt series. Boundary padding (is_abrupt_padded=True) was not yet available — detections reflect instantaneous precision at the point of change.

Code

import pytrendy as pt

df = pt.load_data("series_synthetic")
pt.detect_trends(df, date_col="date", value_col="abrupt",
                 method_params=dict(is_abrupt_padded=False))

Noise detection — random noise on a gradual trend

When the input signal carries moderate random noise layered on top of a gradual trend, the algorithm identifies noise segments while still extracting the underlying Up/Down/Flat structure.

Code

import pytrendy as pt

df = pt.load_data("series_synthetic")
pt.detect_trends(df, date_col="date", value_col="gradual-noisy-20")

Noise detection — spikes on a gradual trend

Isolated outlier spikes sitting on top of a smooth gradual trend are classified as short Noise segments, leaving the surrounding Up/Down structure intact.

This was the initial implementation of spike detection in v1.0.x. Subsequent versions (v1.1.3, v1.1.4, v1.1.7) targeted significant precision improvements — reducing over-wide Noise bands and better distinguishing spike outliers from sustained noise regions.

Code

import pytrendy as pt

df = pt.load_data("series_synthetic")
df.set_index("date", inplace=True)
df.loc["2025-03-01":"2025-03-01", "gradual"] = 500  # inject spike
df = df.reset_index()
pt.detect_trends(df, date_col="date", value_col="gradual")