forecose

PyPI Python versions Tests

A time-series forecasting extension for pydexcom using Google's TimesFM. Readings from the previous 24 hours are captured from the Dexcom Share API service are fed into the model to forecast blood glucose values over the next hour.

All modelling and forecasting is performed locally on your device. The only external connections made are with:

  • Dexcom Share API: fetching CGM readings following the pydexcom approach.
  • HuggingFace: one-time download of the forecasting model weights on the first run.

Quick Start

  1. Ensure that you also the pydexcom package and enabled the Share service within your Dexcom G7 / G6 / G5 / G4.

pip install pydexcom forecose

  1. Initialise pydexcom with your Dexcom credentials (below shows the simplist route, refer to pydexcom for further instruction).
>>> from pydexcom import Dexcom
>>> dexcom = Dexcom(username="username", password="password")
  1. Generate a prediction. By default, the DexcomForecast class predicts the upcoming hour (the next 12 readings) to prevent overextending the forecast. Set custom prediction lengths by adjusting the horizon argument.
>>> from forecose import DexcomForecast
>>> predictions = DexcomForecast(horizon=12).get_forecast(dexcom)
>>> print(predictions)
                          timestamp  predicted_glucose    q10    q25    q50    q75    q90
0  2026-06-30 11:56:43.332500+01:00              156.0  156.0  147.0  155.0  159.0  162.0
1  2026-06-30 12:01:43.332000+01:00              159.0  159.0  142.0  156.0  165.0  169.0
2  2026-06-30 12:06:43.331500+01:00              160.0  159.0  137.0  156.0  169.0  176.0
3  2026-06-30 12:11:43.331000+01:00              160.0  161.0  132.0  155.0  172.0  180.0
4  2026-06-30 12:16:43.330500+01:00              162.0  161.0  128.0  156.0  176.0  185.0
5  2026-06-30 12:21:43.330000+01:00              162.0  162.0  125.0  155.0  177.0  188.0
6  2026-06-30 12:26:43.329500+01:00              163.0  164.0  120.0  155.0  180.0  192.0
7  2026-06-30 12:31:43.329000+01:00              162.0  162.0  116.0  153.0  181.0  194.0
8  2026-06-30 12:36:43.328500+01:00              162.0  162.0  113.0  152.0  183.0  197.0
9  2026-06-30 12:41:43.328000+01:00              161.0  161.0  109.0  151.0  183.0  197.0
10 2026-06-30 12:46:43.327500+01:00              162.0  162.0  107.0  151.0  186.0  201.0
11 2026-06-30 12:51:43.327000+01:00              161.0  160.0  104.0  149.0  186.0  201.0

>>> print(predictions.mmol_l)
                          timestamp  predicted_glucose  q10  q25  q50   q75   q90
0  2026-06-30 11:56:43.332500+01:00                8.7  8.7  8.2  8.6   8.8   9.0
1  2026-06-30 12:01:43.332000+01:00                8.8  8.8  7.9  8.7   9.2   9.4
2  2026-06-30 12:06:43.331500+01:00                8.9  8.8  7.6  8.7   9.4   9.8
3  2026-06-30 12:11:43.331000+01:00                8.9  8.9  7.3  8.6   9.5  10.0
4  2026-06-30 12:16:43.330500+01:00                9.0  8.9  7.1  8.7   9.8  10.3
5  2026-06-30 12:21:43.330000+01:00                9.0  9.0  6.9  8.6   9.8  10.4
6  2026-06-30 12:26:43.329500+01:00                9.0  9.1  6.7  8.6  10.0  10.7
7  2026-06-30 12:31:43.329000+01:00                9.0  9.0  6.4  8.5  10.0  10.8
8  2026-06-30 12:36:43.328500+01:00                9.0  9.0  6.3  8.4  10.2  10.9
9  2026-06-30 12:41:43.328000+01:00                8.9  8.9  6.0  8.4  10.2  10.9
10 2026-06-30 12:46:43.327500+01:00                9.0  9.0  5.9  8.4  10.3  11.2
11 2026-06-30 12:51:43.327000+01:00                8.9  8.9  5.8  8.3  10.3  11.2

What do these predictions mean?

  • predicted-glucose: The most likely trajectory your blood sugar will take (centred baseline of the confidence bands).
  • q10 to q90: The range of confidence bands provide a realistic upper and lower estimate boundaries, showing the full probability distribution of predicted glucose values.

Event Modelling

To account for the key exogenous events (e.g., insulin administration or carbohydrate (carbs) intake) that act on blood glucose values without distorting the underlying TimesFM probability distribution, you can apply a deterministic overlay to your baseline forecast.

Drawing on mathematical frameworks utilised in closed-loop Artifical Pancreas systems and the Hovorka/Bergman meal submodels, event impacts are computed as a second-order linear delay process. Here, event unit rates (e.g., the absorption of insulin or carbs) are translated into a physiological curve that begins slowly, reaches a peak, and then gradually decays over time.

By default, forecose updates the forecast predictions using the standard clinical baselines (a 55-minute peak for insulin, and a 40-minute peak for carbs):

>>> carb_predictions = predictions.add_event(type="carbs", units=30, minutes_ago=0)
>>> print(carb_predictions.mmol_l)
                          timestamp  predicted_glucose   q10  q25   q50   q75   q90
0  2026-06-30 11:56:43.332500+01:00                8.7   8.7  8.2   8.7   8.9   9.0
1  2026-06-30 12:01:43.332000+01:00                9.0   9.0  8.0   8.8   9.3   9.5
2  2026-06-30 12:06:43.331500+01:00                9.3   9.2  8.0   9.0   9.8  10.2
3  2026-06-30 12:11:43.331000+01:00                9.5   9.5  7.9   9.2  10.2  10.6
4  2026-06-30 12:16:43.330500+01:00                9.9   9.8  8.0   9.5  10.7  11.2
5  2026-06-30 12:21:43.330000+01:00               10.2  10.2  8.1   9.8  11.0  11.6
6  2026-06-30 12:26:43.329500+01:00               10.5  10.5  8.1  10.0  11.4  12.1
7  2026-06-30 12:31:43.329000+01:00               10.8  10.8  8.2  10.3  11.8  12.5
8  2026-06-30 12:36:43.328500+01:00               11.0  11.0  8.3  10.5  12.2  13.0
9  2026-06-30 12:41:43.328000+01:00               11.3  11.3  8.4  10.8  12.5  13.3
10 2026-06-30 12:46:43.327500+01:00               11.7  11.7  8.6  11.0  13.0  13.8
11 2026-06-30 12:51:43.327000+01:00               11.9  11.8  8.7  11.2  13.3  14.1

>>> insulin_predictions = predictions.add_event(type="insulin", units=5, minutes_ago=0)
>>> print(insulin_predictions.mmol_l)
                          timestamp  predicted_glucose  q10  q25  q50  q75  q90
0  2026-06-30 11:56:43.332500+01:00                8.6  8.6  8.1  8.5  8.8  8.9
1  2026-06-30 12:01:43.332000+01:00                8.7  8.7  7.7  8.5  9.0  9.2
2  2026-06-30 12:06:43.331500+01:00                8.5  8.5  7.3  8.3  9.0  9.4
3  2026-06-30 12:11:43.331000+01:00                8.3  8.4  6.8  8.0  9.0  9.4
4  2026-06-30 12:16:43.330500+01:00                8.2  8.1  6.3  7.8  8.9  9.4
5  2026-06-30 12:21:43.330000+01:00                7.8  7.8  5.8  7.4  8.7  9.3
6  2026-06-30 12:26:43.329500+01:00                7.5  7.6  5.2  7.1  8.5  9.2
7  2026-06-30 12:31:43.329000+01:00                7.2  7.2  4.6  6.7  8.2  8.9
8  2026-06-30 12:36:43.328500+01:00                6.8  6.8  4.1  6.2  7.9  8.7
9  2026-06-30 12:41:43.328000+01:00                6.4  6.4  3.5  5.8  7.6  8.4
10 2026-06-30 12:46:43.327500+01:00                6.0  6.0  3.0  5.4  7.4  8.2
11 2026-06-30 12:51:43.327000+01:00                5.6  5.6  2.4  4.9  7.0  7.8

By default, sensitivity to insulin (ISF) and carbohydrates (CSF) is set at 40 mg/dL per 1U and 4 mg/dL per gram, respectively. These values are placeholders meant to represent reasonable values and should be adjusted through the add_event method using the sensitivity parameter. In future, I hope to develop a method for calculating estimate values from historic data.

>>> insensitive_insulin_predictions = predictions.add_event(type="insulin", units=5, minutes_ago=0, sensitivity=20.0)
>>> print(insensitive_insulin_predictions.mmol_l)
                          timestamp  predicted_glucose  q10  q25  q50  q75  q90
0  2026-06-30 11:56:43.332500+01:00                8.7  8.7  8.2  8.6  8.8  9.0
1  2026-06-30 12:01:43.332000+01:00                8.8  8.8  7.8  8.6  9.1  9.3
2  2026-06-30 12:06:43.331500+01:00                8.7  8.7  7.4  8.5  9.2  9.6
3  2026-06-30 12:11:43.331000+01:00                8.6  8.7  7.0  8.3  9.3  9.7
4  2026-06-30 12:16:43.330500+01:00                8.5  8.5  6.7  8.2  9.3  9.8
5  2026-06-30 12:21:43.330000+01:00                8.4  8.4  6.4  8.0  9.3  9.9
6  2026-06-30 12:26:43.329500+01:00                8.3  8.4  5.9  7.9  9.3  9.9
7  2026-06-30 12:31:43.329000+01:00                8.0  8.0  5.5  7.5  9.1  9.8
8  2026-06-30 12:36:43.328500+01:00                7.9  7.9  5.2  7.3  9.0  9.8
9  2026-06-30 12:41:43.328000+01:00                7.7  7.7  4.8  7.1  8.9  9.7
10 2026-06-30 12:46:43.327500+01:00                7.5  7.5  4.5  6.9  8.9  9.7
11 2026-06-30 12:51:43.327000+01:00                7.3  7.2  4.1  6.6  8.7  9.5

Key dates

PyPI data
Data sourced directly from PyPI's database.
  • Released:
Newer release available (0.5.0)

1 maintainer

PyPI data
Data sourced directly from PyPI's database.
Avatar for aes21 from gravatar.com aes21

Credits

Author: Alexander Sadler

Requires

Python >=3.10

Report project as malware

Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Source Distribution

forecose-0.4.2.tar.gz (7.6 kB view details)

Uploaded Source

Built Distribution

If you're not sure about the file name format, learn more about wheel file names.

forecose-0.4.2-py3-none-any.whl (7.5 kB view details)

Uploaded Python 3

File details

Details for the file forecose-0.4.2.tar.gz.

File metadata

  • Download URL: forecose-0.4.2.tar.gz
  • Upload date:
  • Size: 7.6 kB
  • Tags: Source
  • Uploaded using Trusted Publishing? Yes
  • Uploaded via: twine/6.1.0 CPython/3.13.12

File hashes

Hashes for forecose-0.4.2.tar.gz
Algorithm Hash digest
SHA256 e9247098190ed7bd6dad33343f7f0bda7520de77a4d0522b1f11a3e192de1943
MD5 836616bd06485c0de71ab0de66611d57
BLAKE2b-256 ce99eb0d2a7a40b643e208b7ad9c49d0b6def5e075fe0ac6ab9b3329df912f6c

See more details on using hashes here.

Provenance

The following attestation bundles were made for forecose-0.4.2.tar.gz:

Publisher: publish.yml on aes21/forecose

Attestations: Values shown here reflect the state when the release was signed and may no longer be current.

File details

Details for the file forecose-0.4.2-py3-none-any.whl.

File metadata

  • Download URL: forecose-0.4.2-py3-none-any.whl
  • Upload date:
  • Size: 7.5 kB
  • Tags: Python 3
  • Uploaded using Trusted Publishing? Yes
  • Uploaded via: twine/6.1.0 CPython/3.13.12

File hashes

Hashes for forecose-0.4.2-py3-none-any.whl
Algorithm Hash digest
SHA256 b82f80442ca85874a5383f717e959baf6cabd9a292b5c4474cf0c4544aef414e
MD5 e0f7f13683c99c63134ecd8299201d83
BLAKE2b-256 3dda1ea8b8d516bc3fd22388ba44c393c9566096c333d4ae78fddf2195009f39

See more details on using hashes here.

Provenance

The following attestation bundles were made for forecose-0.4.2-py3-none-any.whl:

Publisher: publish.yml on aes21/forecose

Attestations: Values shown here reflect the state when the release was signed and may no longer be current.

Release history Release notifications | RSS feed

0.5.0

2 files

0.4.3

2 files

This release

0.4.2 This release

2 files

0.4.1

2 files

0.4.0

2 files

0.3.0

2 files

0.2.2

2 files

0.2.1

2 files

0.2.0

2 files

0.1.1

2 files

0.1.0

2 files

Anthropic, PBC Visionary sponsor Bloomberg Visionary sponsor Hudson River Trading Visionary sponsor Meta Visionary sponsor NVIDIA Visionary sponsor Microsoft Sustainability sponsor Depot Continuous Integration AWS Cloud computing and Security Sponsor Datadog Monitoring Fastly CDN Google Download Analytics Sentry Error logging StatusPage Status page