Pebble rustler

Research engineer @AlloraLabsHQ

[Allora @AlloraNetwork]

New paper from Allora Research: the CZAR loss (Composite Zero-Agnostic Return) for financial return forecasting, led by @PebbleRustler with the Allora Research team.

Under a standard loss like the mean squared error, a model that gets the direction of a return right 58% of the time can score worse than one that always predicts zero. Expected returns are close to zero, so a symmetric loss makes "predict zero" nearly optimal, and models shrink their predictions toward it.

CZAR rewards directional decisiveness instead, with a floor on the loss for small moves so predicting zero can't look artificially good. The tradeoff is a deliberate bias, so its predictions shouldn't be read as expected returns.
It's the scoring loss for every log-return topic on Allora, and it's included in the open-source Forge Builder Kit.

[Apollo11 @apo11o]

I'm excited to announce a new design principle for AI.

Flawed in Nature, Perfect through Evolution: a new paper in Allora Decentralized Intelligence (ADI), showing that intentionally worsening individual models through genetic mutations improves their collective performance. 👇

[Apollo11 @apo11o]

https://t.co/wiRhk1krdj

[Apollo11 @apo11o]

https://t.co/iBEqvGACHz

[Pebble rustler @PebbleRustler]

We're continuously refining Allora's forecasters, the component that predicts which models will perform best under current conditions. Forecaster accuracy is high-leverage as every improvement there flows directly into model weights in the network.

[Apollo11 @apo11o]

https://t.co/TfWSDFisve

[Pebble rustler @PebbleRustler]

Under standard loss functions, a model that predicts zero returns every time can score better than one that's actually trying to predict the market. The new CZAR loss function we've developed fixes this.

Full writeup on the Allora research forum: https://t.co/MjOK6J3ndZ

[Apollo11 @apo11o]

Last week I mentioned a new family of loss functions designed for financial applications. This sounds simple, but it's actually transformative. And we're starting to deploy it on the network now.

The loss function defines what 'accurate' means to the network. It determines each model's score, the score determines its weight, and the weight determines how much that model shapes the final output.

The entire chain from raw inference to network output runs through the loss function. Change it, and the model the network considers its best contributor can flip.

We've seen this clearly in our multi-metric evaluation work. A model that ranks well on magnitude error can rank poorly on directional accuracy. If your loss function only captures magnitude, the network will assign higher weight to models that minimize price distance, but can't tell you whether the market goes up or down. That's a problem if you care about direction.

The family of loss functions we're deploying now was designed by our team to capture directional accuracy alongside magnitude. It changes which models accumulate influence in the network and what the network output looks like in practice.

For a single model, the loss function is a training signal. You optimize against it and move on. For a coordination network, the loss function is a governance decision. It determines which contributors the network rewards, which it sidelines, and what kind of intelligence the end user actually receives.

Two coordination networks with identical participants and identical data will produce different outputs if they use different loss functions.

That's what makes this a high-leverage design choice rather than a technical footnote. The network's intelligence is downstream of what you choose to measure.

[Apollo11 @apo11o]

Last week I mentioned a new family of loss functions designed for financial applications. This sounds simple, but it's actually transformative. And we're starting to deploy it on the network now.

The loss function defines what 'accurate' means to the network. It determines each model's score, the score determines its weight, and the weight determines how much that model shapes the final output.

The entire chain from raw inference to network output runs through the loss function. Change it, and the model the network considers its best contributor can flip.

We've seen this clearly in our multi-metric evaluation work. A model that ranks well on magnitude error can rank poorly on directional accuracy. If your loss function only captures magnitude, the network will assign higher weight to models that minimize price distance, but can't tell you whether the market goes up or down. That's a problem if you care about direction.

The family of loss functions we're deploying now was designed by our team to capture directional accuracy alongside magnitude. It changes which models accumulate influence in the network and what the network output looks like in practice.

For a single model, the loss function is a training signal. You optimize against it and move on. For a coordination network, the loss function is a governance decision. It determines which contributors the network rewards, which it sidelines, and what kind of intelligence the end user actually receives.

Two coordination networks with identical participants and identical data will produce different outputs if they use different loss functions.

That's what makes this a high-leverage design choice rather than a technical footnote. The network's intelligence is downstream of what you choose to measure.

[Allora @AlloraNetwork]

Introducing: ALLO, the world’s first intelligence-backed token.

Let’s explore the utility and tokenomics of ALLO, the native token at the heart of Allora 🧵👇

[Apollo11 @apo11o]

Most ensemble methods weight models based on past performance.

But what if weights could anticipate future accuracy: learning under which conditions each model excels?

That’s the focus of our new paper at @AlloraLabsHQ on context-aware inference via performance forecasting. 👇