v0.2.0 · 31 Tests · 53 Factors Evaluated · 19 Unique Frontier · Interactive Dashboard →

Autonomous Alpha Factor Research
for Chinese A-Shares

AI agents invent, iterate, and optimize quantitative factors guided by Pareto frontier optimization. ~60 experiments per hour, zero human intervention while you sleep.

RankICIC IRTurnover Pareto Front12 Operators

System Architecture

Agent edits
factors.py
Write 1–10 Factor subclasses per experiment with 12 built-in operators
from prepare import Factor, ops
▶
modify & commit
Read-only eval
prepare.py
Auto-discover · 3 metrics · Pareto dominance · archive update
uv run python prepare.py
▶
extract results
Persistent state
pareto_frontier.json
Non-dominated factors · git-tracked · permanent research artifact
factors dominant on ≥1 metric
Three Metrics
First-Principles Evaluation
A factor must predict returns, do so consistently, and be cheap to trade. These three dimensions cannot be maximized simultaneously — the agent discovers the Pareto frontier.
Metric 1 — Predictive Power
RankIC
mean( Spearman(factor, fwd_return) )

Cross-sectional daily Spearman correlation. Range [-1,1]. Uses absolute value for Pareto comparison. Computed over full 2020–2025 period.

Metric 2 — Stability
IC IR
mean(IC) / std(IC)

Information Ratio of daily IC values. Signal-to-noise. IC=0.05 with std=0.10 → IR=0.5 (noisy). IC=0.04 with std=0.04 → IR=1.0 (reliable).

Metric 3 — Tradeability
Turnover
1 − mean(|rank_t − rank_{t−1}|)

Day-over-day ranking stability. 1.0 = zero trading cost (identical ranks daily). 0.5 = half of stocks change rank percentile significantly each day.

Experiment Results
53 Experiments · 48 Factors Evaluated · 0 Crashes
30+ iterations generated 48 factor candidates across 14 categories. All successfully evaluated with zero crashes. The Pareto frontier now contains 19 unique non-dominated factor types (23 entries total). Latest experiment: 10 new factors pushed frontier into high-IC / high-turnover region.
Chart 1 — Pareto Frontier: |RankIC| vs |IC IR| (bubble size = Turnover Stability)
Chart 2 — All 38 Experiments: RankIC Evolution
Chart 3 — Unique Frontier Factors (sorted by |RankIC|)
Pareto Frontier Detail
Factor|RankIC||IC IR|TurnoverCategory

43 total experiments · 29 factors dominated · 8 unique frontier factor types

Write a Factor
3 lines of code. Auto-discovered and evaluated.
The agent writes Factor subclasses in factors.py. Each class is automatically discovered, evaluated against 495 A-shares over 5 years, and checked against the Pareto frontier.
factors.py
from prepare import Factor, ops

class Factor001(Factor):
    name = "momentum_5d"

    def compute(self, df):
        m = df.set_index(["datetime", "symbol"])
        val = ops.cs_rank(m["close"] - ops.delay(m["close"], 5))
        return Factor.as_cs_series(df, val)