From the noise ...

What did I just watch?

TLDR: We crunched and sorted a lot of numbers about the stock market

The videos are two visualizations of the covariance of the Numerai dataset. The dataset contains:

• 20 targets (different values they want you to predict)
• 1050 features (different values you can use to help you make predictions)
• 957 eras (different points in time that represent a 5-business day week)
• 1000s of data records representing different stocks in the stock market, for each era.

The first video shows the covariance data in the order that the features and targets appear in the numerai dataset. The second video shows the data ordered by each feature's average covariance with the 20 targets over time.

That big square? It's a 1050x1050 pixel feature-feature covariance matrix - we use it to help understand the relationships between all the pairs of features (527,043,825 unique covariance calculations) in the numerai dataset.

Those narrow rectangles? Those are a 1050x20 pixel feature-target covariance matrix - we use it to help understand the relationship between all the pairs of features and targets (20,097,000 unique covariance calculations) in the numerai dataset.

That small square? It's a 20x20 pixel target-target covariance matrix - we use it to help understand the relationships between all the pairs of targets (182,590 unique covariance calculations) in the numerai dataset.

Why are you showing me this?

TLDR: It's hard to crunch these numbers, and they help us make better stock market predictions

Calculating large amounts of covariance data like in the case of the Numerai dataset is very resource-intensive - In many situations, people are limited by their technology and can't fully calculate covariance data.

Because calculating statistics on large datasets is so resource-intensive, most ML models built today are black-boxes, which makes it hard to understand how they work, which makes them hard to trust.

With a statistics-first approach, you can build ML models that:

What is Maxia?

TLDR: We're a company working on a platform that makes it faster, cheaper, and easier to build better machine learning models from big time series datasets.

Maxia is a Canadian data and machine learning company based in Toronto. We do research and development on autonomous assembly and optimization of machine learning models for time series data - we're trying to make it easy to build machine learning models just by connecting your data and asking a question. We also work with companies to help them better organize, learn from, and work with their data to improve their products and operations.

In the past, we focused on helping Sales, Marketing, and Customer Support Teams build and use predictions based on customer analytics data from websites and mobile apps. Companies have used our models to predict business objectives like conversion, LTV, and churn to improve their sales, marketing and customer support operations.

In the future, we'll be opening up our machine learning platform more by sharing how it works, and making our technology publicly available again.

Why Numerai?

TLDR: It's a prediction tournament with a public leaderboard - and if we do well, we get rewarded

At Maxia, we've been exclusively researching and building time series machine learning models since our founding in 2019. In 2021, Numerai upgraded its dataset to be much bigger - about the size of other datasets we've normally worked with.

Since October 2021, we've added support for the Numerai dataset to our platform and have been using it exclusively as our research dataset.

The most interesting thing about Numerai is that you can stake cryptocurrency on live predictions that you make, and that staking data is transparent and public. Staking our predictions means that we're confident in our model, and the predictions we're making.

We recently began to stake cryptocurrency on our predictions, and starting in November 2021, Maxia will be increasing its stake.