ts-data-generator 0.4.0

A Python library for generating synthetic time series data

Synthetic Time Series Data Generator

Generate realistic synthetic time series datasets with configurable dimensions, metrics, composable trend functions, and injectable anomalies — via a Python API or the tsdata CLI.

---

Quickstart

CLI (no install)

uvx --python 3.11 --from ts-data-generator tsdata generate \
    --preset daily-sales --output sales.csv

With anomalies and a fixed seed for reproducibility:

uvx --python 3.11 --from ts-data-generator tsdata generate \
    --start 2024-01-01 --end 2024-01-07 --granularity h \
    --dims "region:US,EU,AP" \
    --mets "temperature:SinusoidalTrend(amplitude=10,freq=24)" \
    --anomalies "temperature:PointAnomaly(probability=0.01,magnitude=5)" \
    --seed 42 --output weather.csv

Python API

from ts_data_generator import DataGen
from ts_data_generator.utils.trends import SinusoidalTrend
from ts_data_generator.utils.functions import random_choice
from ts_data_generator.anomalies import PointAnomaly, MissingData

dg = DataGen(seed=42)
dg.start_datetime = "2024-01-01"
dg.end_datetime = "2024-01-07"
dg.to_granularity("h")

dg.add_dimension("region", random_choice(["US", "EU", "AP"]))
dg.add_metric(
    "temperature",
    {SinusoidalTrend(amplitude=10, freq=24)},
    anomalies=[PointAnomaly(probability=0.01, magnitude=5)],
)

print(dg.data.head())
dg.data.to_csv("weather.csv", index_label="datetime")

---

Installation

pip install ts-data-generator

With optional extras:

# Schema imputing (requires scipy)
pip install "ts-data-generator[imputer]"

# Holiday trend support (requires holidays)
pip install "ts-data-generator[holidays]"

# All optional features
pip install "ts-data-generator[all]"

For local development:

git clone https://github.com/manojmanivannan/ts-data-generator.git
cd ts-data-generator
uv sync --extra dev

---

Core Concepts

Dimensions

Categorical or continuous columns generated by an infinite generator function.

Function	Description	CLI shorthand
random_choice	Random element from a collection	name:random_choice:A,B,C
random_int	Random integer in [start, end]	name:random_int:1,100
random_float	Random float in [start, end)	name:random_float:0.0,1.0
constant	Fixed value or cycle	name:constant:10
ordered_choice	Sequential cycle	name:ordered_choice:A,B,C
auto_generate_name	Auto-generated column name	name:auto_generate_name:cat

Shorthand: name:values defaults to random_choice. Example: --dims "product:A,B,C".

Metrics

Numeric columns built by additively composing one or more trends.

Trend	Description	Key parameters
SinusoidalTrend	Sine wave with optional noise	amplitude, freq, phase, noise_level
LinearTrend	Linear ramp with optional noise	limit, offset, noise_level
WeekendTrend	Spikes on Saturday/Sunday	weekend_effect, direction, limit
HolidayTrend	Ramp around holidays	country, effect, pre_window, post_window
ARNoiseTrend	Autoregressive AR(p) noise	coefficients or decay+order, noise_std
MarkovTrend	Discrete-state Markov chain	states, values, stickiness or transition_matrix
StockTrend	Random walk + multi-scale sine	amplitude, direction, noise_level

Trends combine with +: metric_name:Trend1(...)+Trend2(...).

Anomalies

Inject realistic irregularities into metric values. Anomalies are applied per-metric after trend composition and run in order (PointAnomaly → MissingData last, so NaN values are never overwritten).

Anomaly	Description	Key parameters
PointAnomaly	Isolated value spikes	probability, mode (additive/replacement), magnitude
MissingData	NaN gaps	mode (random/burst), probability, min_length, max_length
ConceptDrift	Gradual regime shifts	segments (list of DriftSegment)

PointAnomaly supports two modes:

• additive — adds the magnitude to the trend value at anomalous timestamps.
• replacement — replaces the trend value with the magnitude. Magnitude can be a fixed scalar or a (min, max) tuple for uniform sampling.

MissingData supports three modes:

• random — each timestamp independently becomes NaN with the given probability.
• burst — consecutive blocks of NaN of configurable length, non-overlapping.
• patterned — NaN wherever a schedule callable (pd.Timestamp) -> bool returns True (e.g. every Sunday). Patterned mode composes with random/burst via separate MissingData instances in the anomalies list.

ConceptDrift applies gradual distribution-level shifts using DriftSegment:

from ts_data_generator.anomalies import ConceptDrift, DriftSegment

ConceptDrift(segments=[
    DriftSegment(start_timestamp="2024-01-15T06:00:00",
                 transition_window=1800, target_mean=50, target_std=5,
                 hold_duration=7200, restore=True),
])

Each segment alpha-blends from baseline into N(target_mean, target_std) over transition_window seconds, holds for hold_duration seconds, and optionally transitions back.

Drift positions are specified by absolute start_timestamp. Multi-segment sequences are built by repeating --anomalies for the same metric in the CLI, or by passing a list of segments in the API.

Anomalies combine with + and are scoped to a metric:

metric_name:PointAnomaly(...)+MissingData(...)

Deterministic generation

Pass seed to DataGen(seed=42) or --seed 42 in the CLI for reproducible output. The seed initializes a PCG64-backed SeedableRNG that is threaded through trend generation and anomaly injection, replacing global np.random calls.

Multi-Items

Linked columns generated from a single function — useful when columns have dependencies (e.g. col3 = col1 + col2).

def linked_gen():
    while True:
        a, b = random.randint(1, 100), random.randint(1, 100)
        yield (a, b, a + b)

dg.add_multi_items(names=["val1", "val2", "val3"], function=linked_gen())

Aggregation

Data can be resampled to a coarser granularity with per-metric aggregation methods (sum, mean, min, max):

dg.add_metric("sales", {LinearTrend(limit=50)}, aggregation_type=AggregationType.SUM)
hourly = dg.aggregate("h")  # from 5min -> hourly

---

CLI Reference

tsdata [OPTIONS] COMMAND [ARGS]

generate — create a CSV dataset

tsdata generate \
    --start "2024-01-01" \
    --end "2024-01-31" \
    --granularity "D" \
    --dims "product:A,B,C,D" \
    --dims "region:X,Y,Z" \
    --mets "sales:LinearTrend(limit=1000)+WeekendTrend(weekend_effect=100)" \
    --output "daily_sales.csv"

Options:

Option	Description
--start	Start datetime (YYYY-MM-DD)
--end	End datetime (YYYY-MM-DD)
--granularity	s, min, 5min, h, D, W, ME, Y
--dims	Dimension spec (repeatable)
--mets	Metric spec (repeatable)
--anomalies	Anomaly spec keyed by metric name (repeatable)
--seed	Integer seed for deterministic generation
--output	Output CSV path (must end in .csv)
--preset	Use a built-in preset
--config	Path to a JSON config file

Presets — daily-sales, hourly-metrics, minute-stock, weekly-revenue, monthly-recurring. List with tsdata presets.

Anomaly examples

# Point anomalies
tsdata generate ... --anomalies "sales:PointAnomaly(probability=0.01,magnitude=5)"

# Missing data (random mode)
tsdata generate ... --anomalies "sales:MissingData(probability=0.05)"

# Missing data (burst mode)
tsdata generate ... --anomalies "sales:MissingData(mode=burst,burst_probability=0.02,min_length=3,max_length=10)"

# Missing data (patterned mode — NaN every Sunday)
tsdata generate ... --anomalies "sales:MissingData(mode=patterned,schedule=weekday==6)"

# Concept drift
tsdata generate ... --anomalies "sales:ConceptDrift(start_timestamp=2024-01-15T06:00:00,target_mean=50,target_std=5,hold_duration=7200)"

# Multiple anomaly types on one metric
tsdata generate ... --anomalies "sales:PointAnomaly(probability=0.01,magnitude=5)+MissingData(probability=0.05)"

# Multi-segment concept drift (repeat --anomalies for the same metric)
tsdata generate ... \
    --anomalies "sales:ConceptDrift(start_timestamp=2024-01-01T00:00:00,transition_window=1800,target_mean=50,hold_duration=7200)" \
    --anomalies "sales:ConceptDrift(start_timestamp=2024-01-02T00:00:00,transition_window=3600,target_mean=100,hold_duration=7200,restore=true)"

# Deterministic generation
tsdata generate ... --seed 42

Other commands

tsdata dimensions    # List available dimension functions
tsdata metrics       # List available trend functions
tsdata presets       # List preset configurations
tsdata presets <name>  # Show details for a specific preset

Environment variables

Any option can be set via environment variables prefixed with TSDATA_:

export TSDATA_START="2024-01-01"
export TSDATA_GRANULARITY="h"
tsdata generate --end "2024-01-02" --dims "id:A,B" --mets "val:LinearTrend(limit=10)" --output out.csv

JSON config file

{
  "start": "2024-01-01",
  "end": "2024-01-12",
  "granularity": "5min",
  "dimensions": ["product:A,B,C", "region:X,Y,Z"],
  "metrics": [
    "sales:LinearTrend(limit=500)+WeekendTrend(weekend_effect=50)",
    "orders:LinearTrend(limit=200)"
  ],
  "anomalies": [
    "sales:PointAnomaly(probability=0.01,magnitude=5)+MissingData(probability=0.05)"
  ],
  "seed": 42,
  "output": "data.csv"
}

CLI arguments override config file values.

---

Schema Imputing

Reverse-engineer trend parameters from existing CSV data (requires pip install "ts-data-generator[imputer]"):

from ts_data_generator.schema.converter import SchemaConverter

converter = SchemaConverter("data.csv", index_col=0)
schema = converter.impute_schema()
trends = converter.analyze_numeric_trends(columns=["sales"], top_freq=2)
converter.construct_trend_column("sales", trends["sales"])

See the imputer notebook for a full walkthrough.

---

Example Notebooks

Notebook	Description
sample.ipynb	End-to-end: dimensions, metrics, trends, multi-items, plotting
aggregate.ipynb	Aggregation with multi-items and custom aggregation types
imputer.ipynb	Reverse-engineering schema and trends from existing CSV

---

Package Structure

ts_data_generator/
├── __init__.py            # Public API: DataGen
├── exceptions.py          # Custom exception hierarchy
├── _version.py            # Package version
├── data_gen.py            # DataGen engine (orchestrator)
├── cli.py                 # Click CLI (tsdata command)
├── random.py              # SeedableRNG wrapper (PCG64-backed)
├── anomalies/
│   ├── __init__.py        # Anomaly, PointAnomaly, MissingData, ConceptDrift, DriftSegment
│   ├── base.py            # Abstract Anomaly base class
│   ├── point.py           # PointAnomaly (isolated spikes)
│   ├── missing.py         # MissingData (NaN gaps: random, burst, patterned)
│   └── drift.py           # ConceptDrift + DriftSegment (regime shifts)
├── core/
│   └── dataframe_builder.py  # DataFrame generation logic
├── schema/
│   ├── models.py          # Granularity, AggregationType, Metrics, Dimensions
│   └── converter.py       # CSV schema analysis & trend imputing
├── utils/
│   ├── functions.py       # Dimension generator functions
│   └── trends.py          # Trend generators (Sine, Linear, Weekend, Holiday, ARNoise, Markov, Stock)
└── transforms/
    └── normalizer.py      # Min-max & standard normalization strategies

---

License

MIT — see LICENSE.