A brand new age for time collection

Google simply entered the race of basis fashions for time-series forecasting.

In August 2023, the time-series neighborhood was disrupted by the discharge of TimeGPT, Nixtla’s first basis mannequin for time collection forecasting.

Following TimeGPT, a number of basis forecasting fashions have been launched, however there was one which stood out. Just lately, Google unveiled TimesFM[1], a groundbreaking time-series mannequin with phenomenal outcomes.

Time collection are ubiquitous, utilized in many domains like retail, power demand, economics, healthcare and extra. A basis TS mannequin may be readily utilized to any TS case with nice accuracy, like GPT-4 for textual content.

On this article, we focus on:

1. The challenges of basis fashions in time collection in comparison with NLP.
2. How TimesFM overcomes these challenges.
3. How TimesFM works and why it’s a robust mannequin.
4. TimesFM benchmark outcomes.
5. Prospects for the way forward for basis fashions in time-series forecasting

Let’s get began.

I’ve launched AI Horizon Forecast, a publication specializing in time-series and revolutionary AI analysis. Subscribe right here to broaden your horizons!

The idea of a promising basis mannequin in NLP was already evident with the discharge of GPT-2 in Language Fashions are Unsupervised Multitask Learners [2].

However in time collection, constructing a basis mannequin isn’t easy. There are a number of challenges:

• Dataset Shortage: In NLP, discovering textual content knowledge is simple. Nonetheless, public time-series datasets should not available.
• Unpredictable Format: Language fashions are primarily based on well-defined grammars and vocabularies. Time-series knowledge might belong to domains with totally different traits — e.g., extremely sparse gross sales or unstable monetary knowledge.
• Totally different Granularities: Every time collection mannequin works for a selected granularity — e.g., hourly, weekly, month-to-month…