A-SHARE · MULTI-FACTOR · QP
Explainable, source-auditable research snapshots built from real A-share market data.
Review its method, set parameters, and run the cloud analysis.
RMA FACTOR ENGINE
The engine absorbs evidence, plans hypotheses, generates factor code, and learns from backtest feedback. Every signal remains traceable to a report and a financial mechanism.
REPORT → MEMORY
RMA reads local reports and external research, breaks the text into evidence chunks, screens each chunk for daily OHLCV feasibility, and writes accepted mechanisms into traceable A/B/C memory.
MEMORY → PLAN
Macro Brain combines RMA coverage with GOOD/BAD discovery feedback. It selects one primary mechanism, supporting mechanisms, and report-grounded Research Archetypes for the next cycle.
PLAN → FACTOR
Micro Brain converts each hypothesis into executable Python factor code, then evolves candidates through mutation, crossover, refinement, and novelty injection under strict quality gates.
Operating loop: RMA classifies report evidence → Macro Brain proposes falsifiable hypotheses → Micro Brain writes approved-field factor code and passes the quality gates → candidates enter cost-aware multi-index backtests. Rank IC, alpha, drawdown, and turnover are written back to Cross Brain for the next cycle.
POSITIVE RETURN FACTORS
The filter is positive annualized net return after costs. C801 is the only factor in the archived backtest reports that currently passes this filter across the available universes.
Downside Volatility Bias
The factor measures the asymmetry between downside return shocks and typical volatility, then reverses the cross-sectional signal. The v2 test runs from 2021-01-04 to 2026-08-12, selects the top 50 every five trading days, and includes commissions, stamp duty, and slippage.
NET ASSET VALUE
All series start at 1.00 and show the complete daily backtest path.
DRAWDOWN
Lower values indicate a deeper decline from the previous portfolio peak.
Loading archived backtest series…
| Factor | Universe | Constituent Method | Annual Return | Annual Alpha | Max Drawdown | Rank IC | Status |
|---|---|---|---|---|---|---|---|
| C801 v2 | CSI 300 | Point-in-time constituents | +8.19% | +10.32% | −14.54% | 0.0462 | Positive return · Positive alpha |
| C801 v2 | CSI 500 | Point-in-time constituents | +5.38% | +0.91% | −11.53% | 0.0498 | Positive return · Positive alpha |
| C801 v2 | CSI 1000 | Point-in-time constituents | +6.98% | +3.62% | −9.77% | 0.0522 | Positive return · Positive alpha |
| C801 v2 | Microcap 400 | Current static constituents | +8.38% | +4.98% | −39.80% | 0.0803 | Positive return · Sensitivity test |
Figures come from the archived multi-index reports. The Microcap 400 test backfills current static constituents and is exposed to survivorship bias. C101 and C102 have positive relative alpha in CSI 300 but negative absolute annualized returns, so they are excluded here.
XALPHA · TABLE 1
The paper defines a three-stage transformation from report evidence to OHLCV eligibility, mechanism families, and retrievable Research Archetypes. A C-layer archetype is a research cue, not a factor formula.
| Layer | Input unit | Output | Role in XALPHA |
|---|---|---|---|
| A-layer | Report fragment or evidence chunk | KEEP/DROP decision with an OHLCV feasibility reason | Filters evidence according to the daily OHLCV factor contract. |
| B-layer | A-layer-approved research-path atom | Broad mechanism-family assignment and reusable research path | Groups retained evidence into coarse financial mechanisms for Macro Brain routing. |
| C-layer | B-layer research path | Research Archetype record with mechanism role and report-grounded paths | Constructs actionable archetype memory for hypothesis planning; it is not itself a factor formula. |
Source: XALPHA, Table 1 — Definition of the A/B/C taxonomy used by RMA. Read the paper ↗