← Selected workEirik Otis
INDEPENDENT QUANTITATIVE RESEARCH PROJECT

Autonomous quantitative research system

Bittensor AutoResearch

Creator & System Designer

An end-to-end system for researching whether machine-learning signals could predict relative performance across Bittensor subnet markets—and whether those signals survived realistic portfolio construction and unseen data.

01

THE PROJECT

Building the research infrastructure, not just the model

The project began with a quantitative trading question but developed into a broader research system. It needed to ingest and clean changing market data, engineer comparable features across subnet markets, train and evaluate models, simulate portfolio decisions and preserve strict boundaries between research data and unseen evaluation periods.

The objective was not to find the most attractive historical backtest. It was to build a process capable of testing, rejecting and improving ideas reproducibly.

OBJECTIVE

The objective was to create a repeatable system that could test market hypotheses, evaluate whether apparent predictive signals survived unseen data and translate model outputs into realistic portfolio decisions.

END-TO-END OWNERSHIP

Defining the research objective

Designing the architecture and evaluation process

Directing AI-assisted implementation

Working with Python, data pipelines and model evaluation

Selecting and comparing research approaches

Reviewing outputs and challenging invalid assumptions

Determining whether the system justified capital deployment

02

WHAT I BUILT

Four connected layers

A research system is only as credible as the infrastructure, evaluation and decision policy around the model.

01

Data infrastructure

A repeatable market-data foundation that could survive a changing network.

  • Canonical market datasets
  • Cleaning and validation
  • Repeatable data refreshes
  • Subnet identity and lifecycle handling
  • Feature-generation pipelines
02

Machine-learning research

Comparable experiments focused on ranking relative market performance.

  • Ridge regression
  • XGBoost
  • Cross-sectional market ranking
  • Multiple targets and feature groups
  • Out-of-sample model evaluation
03

Portfolio simulation

A policy layer that translated model output into capital decisions.

  • Candidate ranking
  • Position sizing
  • Liquidity constraints
  • Trading costs
  • AMM-based entry and exit simulation
  • Residual capital held in TAO
04

Recursive research workflow

Controlled iteration without outsourcing evaluation or judgement.

  • Bounded LLM proposals and implementation
  • Deterministic tests and evaluation
  • Repeatable experiment outputs
  • Human review of assumptions and results
I built the environment in which models could be proposed, tested, rejected and audited—not merely one model producing one forecast.
03

HOW THE SYSTEM WORKED

Recursive research within defined boundaries

Automation accelerated the loop. Objectives, constraints and deployment judgement remained human responsibilities.

SYSTEM SCHEMATIC

Human judgement remained outside the automated loop.

Market data, feature generation, modelling, evaluation and portfolio simulation were repeatable system responsibilities. Research objectives, invalid assumptions and deployment decisions remained human ones.

Human

Objectives, architecture and judgement

LLM

Research proposals and implementation assistance

System

Tests, metrics, simulations and reproducible outputs

04

VISUALISING THE RESEARCH

Signal had to survive time and execution

Historical fit, unseen-data performance and real portfolio mechanics were evaluated as separate questions.

EVALUATION / 01Walk-forward evaluation design

The model was repeatedly trained only on information available at the time, then evaluated on later periods it had not seen.

CAMPAIGN / 02Research campaign breadth
231Engineered features
1 · 3 · 7 · 14Target horizons, days
2Model families
100sDaily prediction origins
BOUNDED EXPERIMENTSSELECTED FOR REVIEWFROZEN BOUNDARY

The campaign explored a broad, controlled space. It did not imply that every feature, horizon or configuration was retained.

ATTRIBUTION / 03Signal-to-portfolio attribution
01Signal qualityForecast
02Candidate selectionRanking
03Position sizingExposure
04LiquidityCapacity
05Execution costsRealisation
06Lifecycle accountingContinuity
07Market regimeStability

A model forecast was only one component of the final investment outcome. Portfolio decisions could succeed or fail at every downstream layer.

05

RESULT AND JUDGEMENT

Promising research. Insufficient stability for deployment.

Development results initially suggested that useful predictive relationships might exist. However, the underlying Bittensor network and its subnet markets were changing quickly, and performance did not generalise with enough consistency to justify deploying capital through the strategy.

The durable result was the infrastructure: a reusable system for ingesting data, testing models, simulating portfolio policies and determining when an apparently strong result was not reliable enough.

CONCLUSIONKnowing when not to deploy was part of the result.
06

WHAT IT DEMONSTRATES

System building with judgement

01

Technical systems building

Turning a broad research idea into functioning infrastructure.

02

Quantitative judgement

Separating model performance, portfolio construction and market execution.

03

AI-assisted development

Using LLMs to accelerate implementation without outsourcing evaluation or judgement.

04

Research discipline

Testing promising results against unseen data before committing capital.

The project combined product architecture, quantitative research, technical implementation and practical investment judgement inside one independently built system.