Forecasting enthusiasts & data scientists exploring the art and science of predicting future trends and outcomes.
How this post is connected to other assets
One of the key features that sets AutoTS apart from other forecasting solutions is its model template system. This system allows for the export and import of model configurations, which can streamline the forecasting process in various scenarios.
The code to accomplish this is just these few simple lines, yet the potential it unlocks is significant. We'll explore that idea in the next section.
Train AutoTS on your data and save the best models as a template:
# After fitting an AutoTS model, save the best models
example_filename = "example_export.csv" # .csv/.json
model.export_template(example_filename, models='best',
n=15, max_per_model_class=3)
This export will includes details such as the best selected models, their parameters, and the preprocessing steps applied. Save it as a CSV or JSON file.
As time goes on and new data becomes available, models should be evolved to adapt to any changing dynamics in the time series:
# When you want to train again, evolve from the template
model = AutoTS(forecast_length=forecast_length,
frequency='infer', max_generations=0,
num_validations=0, verbose=0)
model = model.import_template(example_filename, method='only') # method='add on'
The code is simple, but the second and third-order effects of this functionality are more profound than one would expect.
On the initial run of AutoTS, the algorithm is exploring a massive domain of models and transformations to apply to the data.
There are an estimated
7.0 x 10^23
combinations. That's 7 times more options than there are grains of sand on Earth.
You start with randomness, but over time as more generations are run, the algorithm narrows in the better models for your data.
We've seen organizations run the initial search for hundreds of hours. This step only needs to happen once, where each successive time after is an evolution of an already good starting point saved by the model template.
As time series evolve, different models in the template may become more or less effective. The template system allows for easy switching between models as the underlying patterns in the data change.
The other benefits the template system introduces:
Computational efficiency: Starting from a pre-optimized template can significantly reduce computation time compared to starting a full model search from scratch. This can be particularly beneficial when working with large datasets or when frequent forecast updates are required.
Fine-tuning: While templates provide a starting point, AutoTS can still adjust and fine-tune these models based on new data. This combines the benefits of using proven model configurations with the adaptability to changing conditions.
Knowledge transfer: Templates can be shared between different forecasting tasks or even different organizations, allowing for the transfer of knowledge about effective modeling strategies for particular types of time series.
Version control: The ability to export and import templates facilitates version control of modeling strategies. This can be crucial for maintaining consistency and traceability in forecasting processes, especially in regulated industries.
It's important to note that while templates can provide significant benefits, a template that worked well on one dataset may not be optimal for another, even if the datasets seem similar. Therefore, it's often wise to use imported templates as a starting point for further optimization rather than as a final solution.
The model template system in AutoTS represents more than just a convenient feature - it's a powerful tool for capturing and transferring knowledge about time series dynamics across different forecasting tasks and even between organizations.
To understand the full potential of model templates, it's helpful to think of forecasting models as simulators of the underlying dynamics that generate time series data. Each model configuration is essentially attempting to replicate the complex system that produces the observed patterns in the data.
For instance, a model template that successfully forecasts retail sales might incorporate elements that simulate seasonal buying patterns, responses to promotions, and the impact of economic conditions. The power of this approach becomes evident when we consider that similar types of time series often share underlying dynamics.
Sales of Product A at Company X are likely to have similar characteristics to sales of Product A at Company Y. Both might exhibit seasonal fluctuations, day-of-week effects, responses to promotional activities, and sensitivity to broader economic indicators. A model template that effectively captures these dynamics for one company could serve as an excellent starting point for another, even if the specific numbers differ.
This similarity in underlying dynamics across related time series opens up significant opportunities for businesses:
Efficiency and cost savings: Companies can leverage templates developed for one product line or business unit across others, saving time and resources in model development.
Knowledge transfer: Expertise gained in forecasting one type of time series can be easily transferred to similar forecasting tasks, even across different departments or organizations.
Rapid deployment: New forecasting projects can get off to a strong start by beginning with a template that's already proven effective for similar data.
Continuous improvement: As templates are applied and refined across various contexts, they can evolve to capture increasingly complex and nuanced dynamics.
Given the value encapsulated in effective model templates, we envision the role Ouro has in the emergence of a marketplace where companies create and sell these templates to each other.
One powerful aspect of Ouro is its potential for monetization. Users can upload and sell their templates and datasets, creating new revenue streams from their forecasting expertise. This feature not only rewards innovation but also incentivizes the sharing of high-quality, effective forecasting solutions.
This kind of sharing allows businesses and researchers to tap into a global pool of forecasting knowledge, accelerating their ability to develop accurate and robust forecasting models. Whether you're looking to enhance your company's demand forecasting, optimize resource allocation, or tackle complex financial predictions, Ouro provides a valuable resource for accessing proven AutoTS model templates and relevant datasets.
We'll get things started with a template for CPG sales forecasting:
This template represents ~500 of the best models in a search through a subset of the some 7x10^23 possible model configurations. While there is no guarantee these are the very best, they represent a good starting point to evolve from when using your own sales data.
As the Ouro community grows, it has the potential to advance the field of time series forecasting, making sophisticated predictive capabilities more accessible to a wider range of organizations and cultivating a collaborative approach to solving complex forecasting challenges.
To get started, users are encouraged to start by searching for templates relevant to their specific forecasting challenges. These templates can serve as a strong foundation, which users can then adapt and evolve using AutoTS to fit their unique needs. After refining a template, users are encouraged to contribute back to the community by uploading their improved versions, therefore enriching the collective knowledge base. The platform also provides opportunities for users to engage with the forecasting community, sharing insights and collaborating on challenging problems in the #forecasting team.
Discover other posts like this one