Chronosx Quant

A Python library for trading calendar management, execution profiling, and temporal backtesting (Time Travel).

Installation

Install the library:

uv build
pip install dist/chronosx_quant-*-py3-none-any.whl

# install from pypi
pip install chronosx-quant

# `cxq` stands for ChronosX Quant. It is the short command for
# `chronosx-preview`; both commands provide the same functionality.
# shortest forms: q=check, h=holidays, ls=calendars
# calendar aliases: sse, cme, f23, f01, f02
cxq ls
cxq q "2026-08-10 21:00" -c f23
cxq h -c sse -s 2026-08-10 -d 10

# full command:
# list holidays (the default command, SSE by default)
chronosx-preview
chronosx-preview --start 2026-08-10 --days 60

# show the stable numeric calendar IDs
chronosx-preview calendars

# list holidays for another calendar (3 = CN_FUTURES_2300)
chronosx-preview holidays -c 3 --start 2026-08-10

# inspect one timestamp
chronosx-preview check "2026-08-10 21:00" -c 3

# optionally control the scheduler's preloaded window
chronosx-preview check "2026-08-10 21:00" -c 3 \
  --scheduler-start 2026-01-01 --scheduler-end 2027-01-01

Install the extra dependencies for the HTTP service:

uv sync --group docker

Install the default development dependencies for tests and benchmarks:

uv sync

Usage

from chronosx_quant import ChronoTime, use_calendar
import pandas as pd

# use CALENDAR_NAME to select default calendar, e.g. SSE
# use SCHEDULE_START / SCHEDULE_END to control the preloaded schedule window
# defaults: SCHEDULE_START=2022-01-01, SCHEDULE_END=now+3y
time = ChronoTime.now()
time = ChronoTime("2026-03-09 11:29:00+08:00")

# Temporarily use another calendar; the previous one is restored when the
# block exits. Optional start/end arguments control its preloaded window.
with use_calendar("CME Globex Crypto") as scheduler:
    time = ChronoTime("2026-03-09 17:00:00")
    print(scheduler.calendar.name)
    print(time.is_trading())

# time about trading, only support 1min step now
time.is_trading()
# move 2 steps forward(2min), auto skip breaks and weekends
# e.g. 2026-03-09 11:29:00+08:00" -> "2026-03-09 13:01:00+08:00"
time.shift(2)

# shift preserve second and microsecond
time = ChronoTime("2026-03-09 11:29:33.123456+08:00")
# 2026-03-09 11:29:33.123456+08:00" -> "2026-03-09 13:01:33.123456+08:00"
time.shift(2)

# select valid trading times from self to end
# return series of 2 items, 11:29:00 and 13:00:00
time.trading_times(end=pd.Timestamp("2026-03-09 13:01:00+08:00"))
# series can aggregate, e.g. get all date in trading series
time.trading_times(end=pd.Timestamp("2026-03-09 13:01:00+08:00")).resample('D').first()

# simpler trading day calculation
print(ChronoTime("2026-04-08").trading_day_delta("2026-04-09"))

# move to the beginning of trading session which the time belongs to
# e.g. SSE "2026-03-08 11:29:00+08:00" belongs to session '2026-03-08', so the session start is '2026-03-08 09:30:00+08:00'
# e.g. CME session '2026-03-08' starts from '2026-03-07 17:00:00-06:00', so the session start is '2026-03-07 17:00:00-06:00', not '2026-03-08 00:00:00+00:00'
time.to_session_start()

# performance profiling
from chronosx_quant.performance import performance, PerformanceRegistry
@performance("slug_name")
def f1():
    ...
f1()
# get report of this function
print(PerformanceRegistry.get_report("slug_name"))
# get report of all functions
print(PerformanceRegistry.full_report())
# if you want to reset
PerformanceRegistry.clear()

# time travel
from chronosx_quant.mock import travel
with travel("2026-03-09 11:29:00+08:00"):
    # only effect ChronoTime, datetime or pd.Timestamp still work
    # thread-local mock, thread-safe
    ChronoTime.now()

Add calendar

Chronosx based on pandas_market_calendars, so it can use all calendars in the project, and support to add custom calendars.

Project custom calendars live in chronosx_quant/calendars.

For the China futures night-session calendars below, we intentionally diverge from the original upstream calendar model.

The original pandas_market_calendars calendars do not support our multi-break use case cleanly enough, so Chronosx forcefully extends the market-time map with extra open/close events such as break_start_1, break_end_1, break_start_2, break_end_2, break_start_3, and break_end_3. These calendars are meant to be consumed by Chronosx's custom scheduler path, especially StaticMinuteScheduler, which scans our custom open_close_map and builds trading intervals from it.

Because of that, these are Chronosx-specific calendars. Please follow the Chronosx calling pattern when using them:

• use them through ChronoTime, SchedulerManager, StaticMinuteScheduler, or the service API
• do not assume they are interchangeable with upstream built-in calendars in generic pandas_market_calendars workflows
• do not reuse the original single-break calendar assumptions when extending these calendars
• if you add more China futures calendars, keep the custom event naming and Chronosx scheduler contract consistent

For SHF/DCE in China, calendars have multiple breaks. These three built-in variants are available:

• CN_FUTURES_0230 aliases: SC.INE, AG.SHF session: previous day 21:00 to trading day 15:00 hours: 21:00-02:30 | 09:00-10:15 | 10:30-11:30 | 13:30-15:00

• CN_FUTURES_0100 aliases: BC.INE, CU.SHF session: previous day 21:00 to trading day 15:00 hours: 21:00-01:00 | 09:00-10:15 | 10:30-11:30 | 13:30-15:00

• CN_FUTURES_2300 aliases: DCE, CZC session: previous day 21:00 to trading day 15:00 hours: 21:00-23:00 | 09:00-10:15 | 10:30-11:30 | 13:30-15:00

Add scheduler

I use static minute scheduler for speed, don't support multi step in the same time, and don't support extend schedule time range. It's ok to add new scheduler to support multi step or dynamic time range.

• dynamic range scheduler cache?

Benchmark

The benchmark suite uses pytest-benchmark.

Run the full benchmark file:

uv run pytest tests/benchmark_chrono.py --benchmark-only \
  --benchmark-warmup=on \
  --benchmark-calibration-precision=100

Run a single benchmark:

uv run pytest tests/benchmark_chrono.py -k test_perf_is_trading \
  --benchmark-only \
  --benchmark-warmup=on \
  --benchmark-calibration-precision=100

Export the full result for later analysis:

uv run pytest tests/benchmark_chrono.py --benchmark-only \
  --benchmark-warmup=on \
  --benchmark-calibration-precision=100 \
  --benchmark-json=.benchmarks/chrono.json

Run and export only the schedule-size cases:

uv run pytest tests/benchmark_chrono.py -k by_schedule_size \
  --benchmark-only \
  --benchmark-warmup=on \
  --benchmark-calibration-precision=100 \
  --benchmark-json=.benchmarks/schedule-size.json

Useful notes:

• tests/benchmark_chrono.py runs benchmarks across SSE, CME Globex Crypto, ICE, and CN_FUTURES_2300, sampling trading times near the early, middle, and late portions of each loaded schedule
• --benchmark-only runs only benchmark tests and skips normal tests
• warmup and higher calibration precision reduce startup and timer-calibration noise, but benchmarks should still be repeated on an otherwise idle machine
• if you want the usual pytest output without benchmark filtering, you can run uv run pytest tests/benchmark_chrono.py

Benchmark preview:

Source: .benchmarks/chrono.json, generated on 2026-08-14 with chronosx-quant 0.3.0b3, Windows 10, an Intel Core i9-14900HX (32 logical CPUs), and CPython 3.10.19.

Each parameterized benchmark case has its own median. The representative value below is the median of those case medians; the range shows the lowest and highest case median in the file. This is more robust than reporting raw maximum samples, but the wide ranges still show that Windows scheduling affected some cases during this run.

Schedule-size preview:

Source: .benchmarks/schedule-size.json, generated on 2026-08-21 with the schedule-size command above on the same host and Python version. These are warm-state lookup measurements; scheduler construction is not included.

Each cell uses the median of the early, middle, and late query-position case medians for that window. The three lookup methods share the same O(log N) session search. Differences this small are normal measurement noise and should not be interpreted as a schedule-size performance trend; memory use, by contrast, grows approximately linearly with the loaded window.

Docker Service

The container service is implemented with FastAPI and exposes a JSON query API plus a Prometheus-compatible metrics endpoint.

Run locally without Docker:

uv run --group docker -m docker.service

Build and run with Docker:

# build docker image
docker build -t chronosx-quant .
# run container
docker run --rm -p 8000:8000 -e CALENDAR_NAME=SSE chronosx-quant
docker run --rm -p 8000:8000 -e CALENDAR_NAME=SSE -e SCHEDULE_START=2022-01-01 -e SCHEDULE_END=2030-12-31 chronosx-quant

Service schedule window:

• SCHEDULE_START defaults to 2022-01-01
• SCHEDULE_END defaults to now + 3 years
• if you set SCHEDULE_END, use str in any format which pandas.Timestamp(...) can parse, for example 2030-12-31

Health check:

curl "http://localhost:8000/health"

Query the current trading status:

curl "http://localhost:8000/query"

Query a specific time:

curl "http://localhost:8000/query?time=2026-03-10T11:29:00"
curl "http://localhost:8000/query?time=2026-03-10T12:00:00&calendar_name=SSE"

The JSON response includes:

• server_version
• calendar_name
• timezone
• query_time
• is_trading_day
• is_trading_time
• session_start
• session_end
• previous_trading_time
• next_trading_time

Calendar preview:

curl "http://localhost:8000/calendar_preview"
curl "http://localhost:8000/calendar_preview?calendar_name=SSE&days_ahead=32"
curl "http://localhost:8000/calendar_preview?calendar_name=SSE&check_time=2026-08-10%2009:30"

The preview response helps verify upcoming holidays and recent holiday definitions for a calendar. It includes:

• calendar_name
• calendar_full_name
• today
• range_start
• days_ahead
• range_end
• latest_holidays
• upcoming_holidays

When check_time is supplied, check also includes the calendar-local time, date, weekday, whether the date is a trading day, whether the instant is a trading time, and the session's trading_date (or null outside a session).

Prometheus metrics:

curl "http://localhost:8000/metrics"

Example output:

# HELP chronosx_service_info Static service metadata.
# TYPE chronosx_service_info gauge
chronosx_service_info{calendar_name="SSE",timezone="Asia/Shanghai",server_version="chronosx-quant/0.2.2"} 1
# HELP chronosx_trading_day Whether the evaluated time falls on a trading day.
# TYPE chronosx_trading_day gauge
chronosx_trading_day{calendar_name="SSE",timezone="Asia/Shanghai"} 1
# HELP chronosx_trading_time Whether the evaluated time falls inside trading hours.
# TYPE chronosx_trading_time gauge
chronosx_trading_time{calendar_name="SSE",timezone="Asia/Shanghai"} 0

The metrics response is generated with prometheus_client and a per-request custom registry. It avoids global collector state and does not expose query_time as a label.

You can scrape /metrics from Prometheus and alert with:

• chronosx_trading_day == 1 when alerts should only run on trading days.
• chronosx_trading_time == 1 when alerts must be active only during market hours.

Performance Design

Chronosx uses HdrHistogram as the profiling backend for performance. The main reason is predictability: profiling should add consistently low recording overhead, with memory usage that can be determined from the configured range and precision.

Execution latency is a good match for this model. It is a positive duration, recorded as integer microseconds, and expected to remain within a configurable range. HdrHistogram can therefore map each observation directly to a latency bucket instead of dynamically maintaining a set of floating-point centroids.

This gives the profiler:

• predictable memory usage after the maximum latency and significant figures are configured
• consistently low-cost recording, which reduces the profiler's effect on the function being measured
• direct support for tail-latency percentiles such as p99, p999, and p9999
• a natural integer-microsecond representation

TDigest is not unsuitable; it solves a more general problem. It approximates the observed distribution with data-dependent centroids, which is especially useful when the value range is unknown or highly dynamic, or when summaries need to be merged across distributed nodes. Chronosx instead prefers fixed, configurable precision: HdrHistogram quantizes values into deterministic latency buckets whose relative error is controlled by significant_figures. This makes the accuracy, memory usage, and recording overhead easier to reason about than an adaptive centroid summary.

HdrHistogram is still an approximate summary; it does not retain every raw latency measurement. Applications that require mathematically exact percentiles must store the raw samples, or maintain exact counts at every required latency resolution, at a substantially higher and potentially unbounded memory cost. The default here favors predictable, explicitly bounded error and low overhead, not a claim of zero-error percentile calculation or universal performance superiority.

Configuration model:

• the default profile remains microsecond-based: 1 us minimum, 60 s maximum, 3 significant figures
• you can override the global default through PerformanceRegistry.configure_default(...)
• you can configure a named metric through PerformanceRegistry.configure(...) or pass the same options directly to @performance(...) and with performance(...)

All ranges use integer microseconds. significant_figures accepts values from 1 to 5; the default 3 is a practical balance between relative precision and memory usage.

from chronosx_quant.performance import PerformanceRegistry, performance

# Change the defaults used by metrics created after this call.
PerformanceRegistry.configure_default(
    min_value_us=1,
    max_value_us=10_000_000,  # 10 seconds
    significant_figures=3,
)

# Configure one named metric before it records its first value.
PerformanceRegistry.configure(
    "database.query",
    min_value_us=10,
    max_value_us=2_000_000,  # 2 seconds
    significant_figures=4,
)

@performance("database.query")
def query_database():
    ...

# Alternatively, keep a metric's configuration next to its decorator.
@performance(
    "strategy.calculate",
    min_value_us=1,
    max_value_us=500_000,
    significant_figures=3,
)
def calculate_strategy():
    ...

# The same options are available for one measured block.
with performance(
    "market_data.update",
    min_value_us=1,
    max_value_us=100_000,
    significant_figures=3,
):
    update_market_data()

Configure each metric before its first measurement. Once its histogram backend has been created, changing the registry configuration does not rebuild that backend. PerformanceRegistry.clear() removes all collected measurements and restores the library defaults when a complete reconfiguration is required.

Why performance does not use ContextDecorator:

• Chronosx wants one API that works as both a decorator and a with block, but it does not want to share mutable timing state between those two modes
• a plain ContextDecorator style implementation usually stores start_time on self, which is easy to reason about for a single with block, but is much easier to misuse once the same decorator object is entered by concurrent calls
• in Chronosx, decorated functions keep timing state in the local wrapper call frame, so every invocation gets an independent start_time
• with performance(...) still creates a fresh instance per with statement, so context-manager usage also gets isolated state naturally
• this separation keeps the API simple while avoiding accidental cross-thread or re-entrant state corruption from a shared timing attribute

Thread-safety model:

• metric aggregation is shared globally by name, but per-call timing state is not shared
• decorator mode is safe for concurrent calls because timing lives in local variables inside the wrapper, not on the profiler object
• context-manager mode is safe because each with performance(...) expression instantiates a new profiler object before entering the block
• the design goal is not lock-free mutation of every backend detail, but to avoid the much more common bug where overlapping calls overwrite each other's start_time

Overall, the design is optimized for practical in-process latency profiling: HdrHistogram gives Chronosx efficient microsecond-level percentile tracking, and the custom performance implementation keeps per-call timing state isolated so the same API remains straightforward under concurrent use.