skyshin / fold

🪁 A fast Adaptive Machine Learning library for Time-Series, that lets you build, deploy and update composite models easily. An order of magnitude speed-up, combined with flexibility and rigour.

FOLD

Fast Adaptive Time Series ML Engine
Explore the docs »

The Adaptive ML Engine that lets you build, deploy and update Models easily. An order of magnitude speed-up, combined with flexibility and rigour.

Main Features

• 10x faster Adaptive Backtesting - What does that mean?
• Composite Models made Adaptive - What does that mean?
• Distributed computing - Why is this important?
• Update deployed models (coming in May) - Why is this important?

Installation

• Prerequisites: python >= 3.8 and pip

• Install from pypi:

  pip install fold-core
  

Quickstart

You can quickly train your chosen models and get predictions by running:

from sklearn.ensemble import RandomForestRegressor
from statsforecast.models import ARIMA
from fold import ExpandingWindowSplitter, train_evaluate
from fold.composites import Ensemble
from fold.transformations import OnlyPredictions
from fold.utils.dataset import get_preprocessed_dataset

X, y = get_preprocessed_dataset(
    "weather/historical_hourly_la", target_col="temperature", shorten=1000
)

pipeline = [
    Ensemble(
        [
            RandomForestRegressor(),
            ARIMA(order=(1, 1, 0)),
        ]
    ),
    OnlyPredictions(),
]
splitter = ExpandingWindowSplitter(initial_train_window=0.2, step=0.2)
scorecard, prediction, trained_pipelines = train_evaluate(pipeline, X, y, splitter)

Thinking of using fold? We'd love to hear about your use case and help, please book a free 30-min call with us!

(If you install krisi by running pip install krisi you get an extended report back, rather than a single metric.)

Fold is different

• Adaptive Models and Backtesting at lightning speed.
  → fold allows to simulate and evaluate your models like they would have performed, in reality/when deployed, with clever use of paralellization and design.

• Create composite models: ensembles, hybrids, stacking pipelines, easily.
  → Underutilized, but the easiest, fastest way to increase performance of your Time Series models.

• Built with Distributed Computing in mind.
  → Deploy your research and development pipelines to a cluster with ray, and use modin to handle out-of-memory datasets (full support for modin is coming in April).

• Bridging the gap between Online and Mini-Batch learning.
  → Mix and match xgboost with ARIMA, in a single pipeline. Boost your model's accuracy by updating them on every timestamp, if desired.

• Update your deployed models, easily, as new data flows in.
  → Real world is not static. Let your models adapt, without the need to re-train from scratch.

Examples, Walkthroughs and Blog Posts

Name	Type	Dataset Type	Docs Link	Colab
⚡️ Core Walkthrough	Walkthrough	Energy	Notebook	Colab
🚄 Speed Comparison of Fold to other libraries	Walkthrough	Weather	Notebook	Colab
📚 Example Collection	Example	Weather & Synthetic	Collection Link	-
🖋️ Why we ended up building an Adaptive ML engine for Time Series	Blog	Public Release Blog Post	Blog post on Applied Exploration	-

Core Features

• Supports both Regression and Classification tasks.
• Online and Mini-batch learning.
• Feature selection and other transformations on an expanding/rolling window basis
• Use any scikit-learn/tabular model natively!
• Use any univariate or sequence models (wrappers provided in fold-wrappers).
• Use any Deep Learning Time Series models (wrappers provided in fold-wrappers).
• Super easy syntax!
• Probabilistic foreacasts (currently, for Classification, full support coming in April).
• Hyperparemeter optimization / Model selection. (coming in early April!)

What is Adaptive Backtesting?

It's like classical Backtesting / Time Series Cross-Validation, plus: Inside a test window, and during deployment, fold provides a way for models to update their parameters or access the last value. Learn more

Our Open-core Time Series Toolkit

If you want to try them out, we'd love to hear about your use case and help, please book a free 30-min call with us!

Explore our Commercial License options here

Contribution

Join our for live discussion!

Submit an issue or reach out to us on info at dream-faster.ai for any inquiries.

Licence & Usage

We want to bring much-needed transparency, speed and rigour to the process of creating Time Series ML pipelines, while also building a sustainable business, that can support the ecosystem in the long-term. Fold's licence is inbetween source-available and a traditional commercial software licence. It requires a paid licence for any commercial use, after the initial, 30 day trial period.

We also want to contribute to open research by giving free access to non-commercial, research use of fold.

Read more

Limitations

• No intermittent time series support, very limited support for missing values.
• No hierarchical time series support.