Ready to supercharge your AI agents?

Install any skill in seconds. Free to start, no credit card required.

Darts · mkurman | DecimalAI

mkurman/darts

NewSkillScoreMaintained

Darts — time series forecasting library by Unit8. Unified API across ARIMA, Prophet, CatBoost, N-BEATS, TFT, TCN, Transformer, and RNN models. Backtesting, probabilistic forecasting, and covariate support.

AI/MLby mkurman

Links this skill to your agents and keeps it current with the author’s updates.

Automatic

Data Analyst pack

1 · One-time — install the CLI (docs ↗)

pip install decimalai

2 · Get the skill — writes .claude/skills/mkurman-darts/SKILL.md

decimalai skills pull mkurman-darts --out .claude/skills/

No CLI? Get the same file — curl · raw file

curl -fsSL --create-dirs -o .claude/skills/mkurman-darts/SKILL.md https://app.decimal.ai/s/mkurman-darts/SKILL.md

Raw file: /s/mkurman-darts/SKILL.md ↗ · machine-readable summary at /s/mkurman-darts.json ↗

In productionGlobal

how this skill behaves on real traffic · last 30 days

No production data for this skill yet — usage numbers appear here once agents start running it.

Anonymized aggregate across every org using this skill — proof from real production use, not a demo.

Does this skill help? · Benchmark: with vs. withoutsign in to run this on your agents

23 test cases from the publisher’s eval.yaml · each run twice on gemini-3.6-flash — with the skill injected, then without · last run 8d ago

No — -100% lift+5% tokens0% turns83% of whole cases pass with skill (19/23)

Test case	Without → With	Effect	Δ tokens	Δ turns
case-19	✓→✗	▼ Worse	17%	0%
case-02	✓→✓	＝ Same ✓	-5%	0%
case-03	✓→✓	＝ Same ✓	41%	0%
case-04	✓→✓	＝ Same ✓	12%	0%
case-13	✓→✓	＝ Same ✓	137%	0%

Showing 5 of 23 cases

SkillSafety

✓ 3 of 3 passed

Code scan15 of 15 checks clean

0 critical · 0 warning · 0 info — scanner v4 — 8/3/2026

live_secretremote_code_executionreverse_shelldata_exfiltrationssrf_metadataagent_snoopinganti_refusaldestructive_commandsobfuscated_unicodeunicode_homoglyphopaque_blobsuspicious_urltool_overreachtrigger_abuseprompt_injection_phrasing

IntentSafe

LLM judge · does the skill do what it claims

ContentClean

LLM judge · policy & harmful-content review

Source

Imported

mkurman / zorai

skills/scientific-skills/darts/SKILL.md

View on GitHub ↗

License MITImported 8/3/2026Version v1

Skill contents

SKILL.md

Overview

Darts (Unit8) provides a unified forecasting API across statistical models (ARIMA, Prophet, Theta), deep learning (N-BEATS, TFT, TCN, Transformer, RNN), and ensemble methods. Supports univariate/multivariate, probabilistic forecasting, covariate handling, and backtesting.

Installation

bash
uv pip install darts

Basic Forecast

python
from darts import TimeSeries
from darts.models import ExponentialSmoothing
import pandas as pd

series = TimeSeries.from_dataframe(pd.DataFrame({"y": [1,2,3,4,5,6,7,8,9,10]}), value_cols="y")
model = ExponentialSmoothing()
model.fit(series)
forecast = model.predict(6)
print(forecast.values())

Deep Learning (N-BEATS)

python
from darts.models import NBEATSModel
model = NBEATSModel(input_chunk_length=24, output_chunk_length=12)
model.fit(train, epochs=100)
pred = model.predict(12)

Backtesting

python
from darts.metrics import mae, mape

errors = model.backtest(series, start=0.7, forecast_horizon=6, stride=1)
print(f"MAE: {mae(errors):.3f}, MAPE: {mape(errors):.3f}")

References

Want to see if mkurman/darts would activate on your phrasing? Try it before you install — no LLM call, no signup.

Latest benchmark · publisher runVerified

8/21/2026, 08:58 UTC · 23 cases

Lift

-100%

0% vs 100%

Agent turns

1.0 vs 1.0

Tokens

+5%

1,451.26 vs 1,378.13

pass → pass · 19fail → fail · 3pass → fail · 1

Per-case results (23)

Case	Status	Duration (ms)			Turns			Tokens			Tool calls
Case	Status	Without	With	Δ	Without	With	Δ	Without	With	Δ	Without	With	Δ
case-01 How do I conduct a historical backtest on a time series using Darts to assess model performance? Please show a Python example that executes a rolling backtest and computes accuracy metrics like MAE and MAPE.	fail→fail	12,711	10,365	-18%	1	1	0%	2,338	2,358	+1%	0	0	—
case-02 I am setting up a virtual environment using uv as my package manager and need to install the Unit8 time series forecasting library. What shell command installs this framework?	pass→pass	5,199	3,084	-41%	1	1	0%	899	851	-5%	0	0	—
case-03 I have a pandas DataFrame with target values stored in column 'y'. I want to construct a native time series object in the Unit8 forecasting framework. Instead of using raw numpy array converters, what static factory method converts the DataFrame directly?	pass→pass	4,105	3,452	-16%	1	1	0%	669	944	+41%	0	0	—
case-04 I want to fit an Exponential Smoothing model on a time series object named `series` using the Unit8 unified forecasting API and predict 6 steps ahead. Base statistical packages use `fit()` and `forecast()`. What model class and method call sequence implements this in the framework?	pass→pass	4,146	3,075	-26%	1	1	0%	812	913	+12%	0	0	—
case-13 From which top-level module in the Unit8 Python forecasting library is the primary `TimeSeries` class imported?	pass→pass	1,690	1,529	-10%	1	1	0%	240	569	+137%	0	0	—
case-05 After running predict on a statistical forecasting model in the Unit8 framework, I receive a prediction time series object. I want to extract its underlying numerical values into an array. Standard pandas uses `.to_numpy()`. What method on the prediction object directly returns the array of values?	pass→pass	5,430	3,429	-37%	1	1	0%	897	978	+9%	0	0	—
case-06 When instantiating `NBEATSModel` in the Unit8 deep learning forecasting suite, I need to set the historical lookback to 24 steps and the forecast horizon to 12 steps. Other neural network packages call these `lookback_window` and `horizon`. What parameter names must be passed to the constructor?	pass→pass	2,568	2,386	-7%	1	1	0%	496	812	+64%	0	0	—
case-07 I am training an N-BEATS deep learning model in the Unit8 framework using a training series named `train`. I want to run training for 100 epochs. Frameworks like Keras or PyTorch Lightning use `max_epochs` or `nb_epochs`. What keyword argument is passed to `.fit()`?	pass→pass	11,306	3,518	-69%	1	1	0%	2,169	1,068	-51%	0	0	—
case-08 When executing `model.backtest()` on a time series in the Unit8 forecasting library, I want to start backtesting at 70% of the series length (0.7), use a forecast window of 6 steps, and advance by 1 step at each iteration. What keyword arguments specify these parameters?	pass→pass	4,295	3,245	-24%	1	1	0%	735	937	+27%	0	0	—
case-09 From which subpackage module within the Unit8 forecasting framework should error evaluation functions like `mae` and `mape` be imported, rather than using `sklearn.metrics`?	pass→pass	3,475	2,689	-23%	1	1	0%	578	634	+10%	0	0	—
case-10 I am using Meta's Prophet library directly (`from prophet import Prophet`) to model time series with custom US national holiday effects. How do I construct a pandas DataFrame containing custom holiday dates and pass it into the `Prophet()` object?	pass→pass	10,014	7,803	-22%	1	1	0%	2,015	1,932	-4%	0	0	—
case-11 I am building a seasonal ARIMA model directly using `statsmodels.tsa.statespace.sarimax.SARIMAX`. How do I set the seasonal order parameter (p, d, q, s) = (1, 1, 1, 12) in the `SARIMAX` constructor?	pass→pass	3,915	3,475	-11%	1	1	0%	770	1,087	+41%	0	0	—
case-12 I am building a raw PyTorch custom `Dataset` and `DataLoader` for multi-variate time-series sliding windows using `torch.utils.data.Dataset`. How do I implement the `__getitem__` method to slice a 2D torch Tensor into input and target windows?	pass→pass	14,495	13,647	-6%	1	1	0%	3,057	3,084	+1%	0	0	—
case-14 Where are model classes like `ExponentialSmoothing` and `NBEATSModel` imported from within the Unit8 forecasting framework package structure?	pass→pass	4,548	4,725	+4%	1	1	0%	824	1,215	+47%	0	0	—
case-15 In the Unit8 time series suite, `model.backtest(...)` yields error evaluation output. How is the `mae` metric function applied to that output to calculate the overall mean absolute error?	fail→fail	13,054	12,273	-6%	1	1	0%	2,285	2,607	+14%	0	0	—
case-16 When generating forecasts from a fitted statistical model in the Unit8 framework, how do you pass the number of future forecast steps (e.g., 6 steps) to the `predict` method?	pass→pass	4,084	3,246	-21%	1	1	0%	534	851	+59%	0	0	—
case-17 What is the exact class name for N-BEATS deep learning model initialization in the Unit8 forecasting suite, and from which submodule is it imported?	pass→pass	4,432	2,781	-37%	1	1	0%	754	840	+11%	0	0	—
case-18 Besides N-BEATS, which deep learning architectures for time series forecasting are supported natively as built-in model classes in the Unit8 package?	pass→pass	8,576	6,611	-23%	1	1	0%	1,558	1,542	-1%	0	0	—
case-19 Which classical statistical forecasting models are provided natively within the Unit8 framework alongside ARIMA?	pass→fail	7,412	6,471	-13%	1	1	0%	1,346	1,576	+17%	0	0	—
case-20 Does the Unit8 Python forecasting framework include built-in capabilities for ensembling multiple underlying forecasting models?	pass→pass	9,120	4,927	-46%	1	1	0%	1,610	1,223	-24%	0	0	—
case-21 In the Unit8 forecasting suite, what features exist for handling multiple variables, uncertainty quantification, and external covariate series?	pass→pass	14,132	14,187	+0%	1	1	0%	2,571	3,015	+17%	0	0	—
case-22 In the Unit8 forecasting suite, how is the `mape` metric function from `darts.metrics` evaluated on the result of `model.backtest(...)`?	fail→fail	13,197	12,435	-6%	1	1	0%	2,575	2,643	+3%	0	0	—
case-23 When building a complete basic forecasting pipeline in the Unit8 framework using a pandas DataFrame, what sequence of operational steps is required from data loading to printing numerical predictions?	pass→pass	10,498	6,801	-35%	1	1	0%	1,965	1,700	-13%	0	0	—

These cases were also screened by the skill’s intent / safety pass.See the Trust review ↗Reproduce this benchmark in Colab ↗

How this skill was measured

DecimalAI ran this skill against gemini-3.6-flash twice over the same eval suite — once with the skill loaded and once without — and compared the two runs case by case. 23 cases were attempted. The headline lift of -100 percentage points is the difference between those two pass rates over the 23 comparable cases. 1 case got worse with the skill loaded, and it is included in that figure.

Model measured: gemini-3.6-flash
Graded by: gemini-3.6-flash
Grading method: judged
Verification: verified
Pass rate without the skill: 100%
Pass rate with the skill: 0%
Cases attempted: 23
Cases counted toward lift: 23
Cases that regressed: 1
Skill version measured: v1
Run completed: 8/21/2026, 08:58 UTC

SkillScore blends 4 signalsno signals yet

BenchmarkRun a verified benchmark with positive lift

Live evalNeeds ≥20 evaluated traces in 30 days

AI ratingNeeds ≥10 rated traces

AdoptionNeeds ≥10 serves across ≥2 orgs in 30 days

Efficiency · separate axis (at equal correctness)+0% turns+5% tokens

Shown for transparency; not credited unless a verified run shows no correctness regression.

No signals yet — run a benchmark or collect evaluated traces to earn a score.

Verified on gemini-3.6-flash · re-tested 8 days ago

Other measured skills in the registry, with their headline benchmark lift.