AlphaDiverse: Post-Training Local Quantitative Research Agents for Diverse Exploration in Alpha Factor Mining

Qingzhuo Wang; Zikun Wei; Zhihua Wei; Wen Shen

cs.AISep 24, 2026

AlphaDiverse: Post-Training Local Quantitative Research Agents for Diverse Exploration in Alpha Factor Mining

Authors: Qingzhuo Wang, Zikun Wei, Zhihua Wei, Wen Shen

Organizations: Tongji University · Shanghai Non-convex Intelligent Technology

Abstract

Large language model (LLM)-based multi-agent systems can automate alpha factor mining, but their reliance on external APIs limits control over cost, availability, and confidentiality. Long research loops also tend to revisit a few successful economic mechanisms that lead to research path collapse. To address these limitations, we propose AlphaDiverse, a framework that integrates a multi-agent alpha research system, diverse research path collection, and post-training for local agents. We let the research system generate complementary plan portfolios and vary research environments across loops to collect diverse research paths. Using these diverse traces, we warm-start local Planner and Realizer agents with supervised fine-tuning. Then, we propose a joint GRPO method to optimize both of them using predictive quality and diversity of contributions. Research feedback is confined to inner period data, while a frozen final model is evaluated on a later outer period data, thereby avoiding test-set tuning. Experiments across four Chinese stock universes show that AlphaDiverse can combine competitive prediction with broader exploration.

Figures & tables

Figure 1: Illustration of the proposed AlphaDiverse framework. API-based research loops produce diverse plans, factor implementations, and evaluation feedback. We select high-quality, complementary traces to warm-start local Planner and Realizer agents, then jointly refine them with GRPO. The trained nodes replace their API counterparts for locally controlled alpha factor research.

Category	Method	Predictive performance				Portfolio performance
Category	Method	IC $\uparrow$	ICIR $\uparrow$	RIC $\uparrow$	RICIR $\uparrow$	ARR $\uparrow$	IR $\uparrow$	MDD $\downarrow$	CR $\uparrow$
	Ridge	0.0166	0.0865	0.0210	0.1184	14.44	0.889	13.49	1.071
	MLP	0.0287	0.1320	0.0377	0.1905	27.34	1.604	13.34	2.050
ML	LightGBM	0.0348	0.2043	0.0335	0.2243	24.15	1.846	6.99	3.456
	GRU	0.0303	0.1165	0.0170	0.0720	18.46	1.024	13.81	1.316
	LSTM	0.0348	0.1460	0.0277	0.1192	24.83	1.280	16.50	1.505

Table 1: Experiment results about CSI300 predictive and portfolio performance. ARR/MDD are percentages; best and second-best values are bold and underlined (lower MDD is better). Other markets and API backends appear in Appendices D.1 and D.2 .

Method	CSI300			CSI500
Method	M $\uparrow$	U $\uparrow$	C $\uparrow$	M $\uparrow$	U $\uparrow$	C $\uparrow$
RD-Agent(Q)	18	3	67	25	14	121
AlphaAgent	5	5	16	5	1	34
QuantaAlpha	11	3	58	22	4	64
AlphaSchema	18	2	96	19	7	96
Single Synthesis	19	8	122	20	10	99

Table 2: Results of diversity on CSI300 and CSI500 across agentic methods.

Variant	RIC $\uparrow$	ARR $\uparrow$	M $\uparrow$	U $\uparrow$	Pair $\downarrow$
- w/o complementarity	0.0352	37.73	10	7	0.455
- w/o retrieval	0.0373	28.39	25	9	0.371
- w/o memory	0.0383	32.26	21	11	0.473
Single Synthesis	0.0316	23.83	19	8	—
AlphaDiverse (GPT-5.5)	0.0398	35.97	24	11	0.322

Table 3: Results of Module ablations (GPT-5.5).

Data selection	RIC $\uparrow$	ARR $\uparrow$	M $\uparrow$	U $\uparrow$	Pair $\downarrow$
- w/ improvement only	0.0371	22.38	21	1	0.616
- w/o balancing	0.0383	23.95	25	3	0.546
- w/o paired supervision	0.0365	21.51	19	1	0.651
full data SFT	0.0389	24.77	27	4	0.511

Table 5: SFT-data ablations.

Appendix figures & tables20 assets

Supplementary material from the paper’s appendix.

Appendix

Object	Example and role
Environment $\mathcal{E}$	$\mathcal{U}$ : historical CSI500 constituents; $y$ : next open-to-open excess return; $\mathcal{V}$ : daily and minute price–volume features; $\mathcal{B}$ : 21 initial factors; $\mathcal{O}$ : registered operators; $\mathcal{G}$ : mechanisms feasible for these features and operators; $\eta$ : inner time folds and later outer evaluation.
Initial state $s_{1}$	$(\mathcal{E},\mathcal{B},\varnothing)$ : the environment, the initial factor set, and no earlier research feedback.
Plan portfolio $\mathcal{P}_{1}$	$p_{1,1}$ : market residual continuation; $p_{1,2}$ : late-session pressure reversal; $p_{1,3}$ : volume–price confirmation; $p_{1,4}$ : range compression.
Expanded $p_{1,4}$	Event: unusually compressed trading range. Mechanism: temporary balance may precede renewed price movement. Horizon: 1 day. Condition: distinguish quiet from fragile states. Direction: nonlinear. Relation to base: a new direction.
Retrieved cards $\mathcal{K}_{1,4}$	Features: high_low_range (daily relative range), minute_open_5m_range (opening-window range). Operators: ts_rank (trailing within-stock percentile rank), at_divide (protected ratio). Transforms: identity , at_reverse (negation).
Specifications $\mathcal{Z}_{1,4}$	Two explicit implementations of the range-compression plan, shown in Figure 5 . Other plans receive their own cards.

Appendix

Table 7: An example of the data passed between research modules. The symbols follow Section 2 .

Theme	Count	Mechanisms
Trend and momentum	5	short horizon momentum; medium horizon momentum; trend strength efficiency; breakout continuation; return acceleration
Reversal	6	short horizon reversal; overreaction exhaustion; opening gap reversal; late session pressure reversal; VWAP dislocation reversal; range oscillation
Liquidity and volume	6	liquidity shock; illiquidity price impact; trading-activity surprise; volume price confirmation; volume price disagreement; liquidity dry up
Intraday structure	6	opening pressure persistence; late session persistence; intraday return asymmetry; intraday volume concentration; high low timing; intraday VWAP pressure
Volatility and range	7	realized volatility regime; range expansion; range compression; volatility of volatility; downside upside asymmetry; tail risk extremes; intraday distribution shape
Relative returns and crowding	8	market relative strength; market beta exposure; idiosyncratic volatility; market residual continuation; market residual reversal; cross sectional dispersion; co movement crowding; dependence regime

Appendix

Table 8: The full economic-mechanism catalogue. The environment-specific

\mathcal{G}

contains its feasible entries.

Criterion	Research acceptance	Realizer pair selection
Factor–target coverage	$\geq 0.20$	$\geq 0.80$
Days with valid IC	$\geq 2$	$\geq 120$
Factor score $s_{\mathrm{fac}}$	$\geq 0.05$	$\geq 0.25$
Absolute RIC	Included in score	$\geq 0.01$
Absolute RICIR	Included in score	$\geq 0.05$
Mean absolute correlation	$\leq 0.90$ against references	$\leq 0.80$ within the pair

Appendix

Table 9: Thresholds for research factors and Realizer pair examples. Both factors in a training pair must pass the right-hand column and have a recorded pair correlation.

Factor	Formula	Factor	Formula
ret_1	$c_{t}/c_{t-1}-1$	open_close_ret	$c/o-1$
high_low_range	$h/l-1$	close_position	$(c-l)/q$
vwap	$w$	vwap_deviation	$c/w-1$
log_turnover	$\log(1+a)$	log_volume	$\log(1+v)$
KMID	$(c-o)/o$	KLEN	$q/o$
KMID2	$(c-o)/q$	KUP	$(h-u)/o$

Appendix

Table 10: All 21 initial factors in

\mathcal{B}

. For one stock and date,

o,h,l,c,v,a

denote OHLC prices, share volume, and traded amount;

w=a/v

q=h-l

u=\max(o,c)

, and

d=\min(o,c)

. Zero denominators give missing values.

Set	Examples	Interpretation
$\mathcal{V}$ , daily	ret_20 , volatility_20 , RSV20	Medium-horizon return, return variation, and position within a recent price range.
$\mathcal{V}$ , minute	last30_ret , first_half_ret , close_vwap_deviation	Closing/earlier-session moves and displacement from intraday VWAP.
$\mathcal{O}$ , arithmetic	at_subtract , at_divide , linear_combo	Differences, protected ratios, and weighted combinations.
$\mathcal{O}$ , temporal	ts_mean , ts_rank	Trailing means and historical percentile ranks within each stock.
$\mathcal{O}$ , transforms	cs_rank , cs_zscore , at_reverse	Same-date ranking, standardization, and sign reversal.

Appendix

Table 11: Examples from available features

\mathcal{V}

and permitted operators/transforms

\mathcal{O}

. Traded amount is monetary value; it is not the share-turnover rate.

Role	Examples	Input tokens	Supervised tokens	Sequence tokens
Planner	589	9,863,796	341,416	10,205,212
Realizer	1,006	2,654,667	402,463	3,057,130

Appendix

Table 12: Final SFT dataset sizes and token counts, including repair examples. Supervised tokens belong to the target assistant responses; sequence tokens are the sum of input and supervised tokens.

Dimension	Group	Source loops	Planner examples	Realizer examples
API backend	GPT-5.5	24	196	253
	Grok-4.6	24	206	359
	GLM-5.3	24	187	394
Market	CSI300	18	111	171
	CSI500	18	130	256
	CSI1000	18	205	311

Appendix

Table 13: SFT examples by source configuration. The four dimension blocks summarize the same datasets and are not additive across blocks.

Role	Responses	Input tokens	Output tokens	Total tokens
Planner	768	6,329,344	517,551	6,846,895
Realizer	1,220	6,088,926	1,023,290	7,112,216
Total	1,988	12,418,270	1,540,841	13,959,111

Appendix

Table 14: Actual GRPO response and token counts across training. These are generation counts, with input and output tokens reported separately.

Field	Instruction
Input	Anonymous ID, executed formula, economic rationale, and the complete mechanism catalogue. All inputs are causal daily/minute price–volume features.
Task	Assign exactly one primary catalogue ID to the predictive economic mechanism. Identify the driver and event, not the factor name or operator syntax.
Evidence	Prioritize the executed formula when it conflicts with the rationale; flag the conflict. A normalization or volatility denominator alone is not a new mechanism.
Boundaries	Window and scale changes preserve the main mechanism. Distinguish persistence from reversal using direction and economic intent. Do not infer profitability or method identity.
Unknown	Use other and define a genuinely new mechanism; use uninterpretable when the available evidence is insufficient. Do not force a match.
Output	Exact anonymous ID, one primary ID, up to two diagnostic secondary IDs, short formula-grounded evidence, conflict flag, and a definition when using other .

Appendix

Table 15: Shared prompt card for independent catalogue annotation.

Category	Method	Predictive performance				Portfolio performance
Category	Method	IC $\uparrow$	ICIR $\uparrow$	RIC $\uparrow$	RICIR $\uparrow$	ARR $\uparrow$	IR $\uparrow$	MDD $\downarrow$	CR $\uparrow$
	Ridge	0.0050	0.0232	0.0225	0.0976	-3.25	-0.156	17.83	-0.182
	MLP	0.0150	0.0734	0.0296	0.1347	15.01	0.810	11.42	1.342
ML	LightGBM	0.0166	0.0910	0.0228	0.1146	-6.74	-0.432	15.65	-0.431
	GRU	0.0201	0.1008	0.0309	0.1567	16.40	0.941	10.55	1.555
	LSTM	0.0164	0.0788	0.0295	0.1380	21.30	1.233	8.40	2.537

Appendix

Table 16: CSI500 predictive and portfolio performance. ARR/MDD are percentages.

Category	Method	Predictive performance				Portfolio performance
Category	Method	IC $\uparrow$	ICIR $\uparrow$	RIC $\uparrow$	RICIR $\uparrow$	ARR $\uparrow$	IR $\uparrow$	MDD $\downarrow$	CR $\uparrow$
	Ridge	0.0015	0.0076	0.0219	0.0998	-16.05	-0.751	20.97	-0.765
	MLP	0.0134	0.0725	0.0230	0.1232	0.09	0.004	21.78	0.004
ML	LightGBM	0.0059	0.0332	0.0226	0.1159	-13.43	-0.654	21.60	-0.622
	GRU	0.0125	0.0660	0.0088	0.0557	-3.24	-0.276	11.78	-0.228
	LSTM	-0.0011	-0.0062	0.0083	0.0575	-19.68	-1.348	20.03	-0.982

Appendix

Table 17: CSI1000 predictive and portfolio performance, January–August 2026. ARR/MDD are percentages. Definitions follow Table 1 .

Category	Method	Predictive performance				Portfolio performance
Category	Method	IC $\uparrow$	ICIR $\uparrow$	RIC $\uparrow$	RICIR $\uparrow$	ARR $\uparrow$	IR $\uparrow$	MDD $\downarrow$	CR $\uparrow$
	Ridge	0.0123	0.0713	0.0409	0.2105	7.51	0.649	8.18	0.918
	MLP	0.0188	0.1146	0.0349	0.1948	9.90	0.986	7.73	1.249
ML	LightGBM	0.0246	0.1781	0.0381	0.2019	2.98	0.275	11.19	0.266
	GRU	0.0147	0.1044	0.0172	0.1072	-1.63	-0.160	11.38	-0.143
	LSTM	0.0145	0.0991	0.0202	0.1251	-4.25	-0.355	14.33	-0.297

Appendix

Table 18: A-share predictive and portfolio performance, January–August 2026. ARR/MDD are percentages. Definitions follow Table 1 .

Backend	IC	ICIR	RIC	RICIR	ARR	IR	MDD	CR	M	U	C
CSI300
GPT-5.5	0.0388	0.1995	0.0398	0.2446	35.97	2.872	7.08	5.083	24	11	108
Grok-4.6	0.0285	0.1540	0.0324	0.1986	21.53	1.503	7.70	2.796	30	17	116
GLM-5.3	0.0373	0.2105	0.0372	0.2418	26.41	2.214	9.24	2.858	31	18	132
CSI500
GPT-5.5	0.0286	0.1979	0.0347	0.2371	6.12	0.426	10.33	0.592	24	11	125

Appendix

Table 19: Predictive performance, portfolio performance, and research diversity of API-based AlphaDiverse across four markets. ARR/MDD are percentages; M, U, and C denote mechanisms, useful mechanisms, and signal clusters, respectively.

Method	CSI300			CSI500			CSI1000			A-share
	M $\uparrow$	U $\uparrow$	C $\uparrow$	M $\uparrow$	U $\uparrow$	C $\uparrow$	M $\uparrow$	U $\uparrow$	C $\uparrow$	M $\uparrow$	U $\uparrow$	C $\uparrow$
RD-Agent(Q)	18	3	67	25	14	121	15	9	127	8	8	23
AlphaAgent	5	5	16	5	1	34	1	1	19	3	3	20
QuantaAlpha	11	3	58	22	4	64	21	4	63	13	5	59
AlphaSchema	18	2	96	19	7	96	19	10	99	18	6	94
Single Synthesis	19	8	122	20	10	99	21	11	100	22	12	101

Appendix

Table 20: Research diversity across four markets. Best and second-best values within each market and metric are bold and underlined; ties receive the same mark.

	Retained mechanisms $U\uparrow$
Method	CSI300					CSI500					CSI1000					A-share
	32	64	96	128	160	32	64	96	128	160	32	64	96	128	160	32	64	96	128	160
GPT-5.5	6	8	8	8	11	6	8	11	11	11	9	14	17	19	19	6	12	12	15	17
Grok-4.6	6	14	14	14	17	8	11	13	13	14	12	15	18	20	21	13	16	19	22	23
GLM-5.3	7	12	13	13	18	5	7	7	7	7	10	17	26	28	28	10	10	16	19	21
Untrained	8	11	11	11	11	0	0	3	3	6	6	8	10	11	11	12	16	17	18	19

Appendix

Table 21: Model-retained mechanism coverage U at five candidate budgets.

Market	Variant	Economic performance								Research diversity
Market	Variant	IC $\uparrow$	ICIR $\uparrow$	RIC $\uparrow$	RICIR $\uparrow$	ARR $\uparrow$	IR $\uparrow$	MDD $\downarrow$	CR $\uparrow$	M $\uparrow$	U $\uparrow$	C $\uparrow$	Pair corr. $\downarrow$
CSI300	- w/o complementarity	0.0352	0.1901	0.0352	0.2284	37.73	3.821	5.03	7.501	10	7	85	0.455
	- w/o retrieval	0.0329	0.1692	0.0373	0.2166	28.39	2.256	11.28	2.517	25	9	116	0.371
	- w/o memory	0.0370	0.1939	0.0383	0.2084	32.26	2.176	5.29	6.100	21	11	92	0.473
	Single Synthesis	0.0319	0.1862	0.0316	0.2037	23.83	1.914	10.10	2.359	19	8	122	—
	AlphaDiverse	0.0388	0.1995	0.0398	0.2446	35.97	2.872	7.08	5.083	24	11	108	0.322

Appendix

Table 22: Module ablations with GPT-5.5 across four markets. ARR/MDD are percentages.

Market	Variant	Economic performance								Research diversity
Market	Variant	IC $\uparrow$	ICIR $\uparrow$	RIC $\uparrow$	RICIR $\uparrow$	ARR $\uparrow$	IR $\uparrow$	MDD $\downarrow$	CR $\uparrow$	M $\uparrow$	U $\uparrow$	C $\uparrow$	Pair corr. $\downarrow$
CSI300	- w/o training	0.0257	0.1432	0.0291	0.1797	22.49	1.720	10.05	2.238	33	11	121	0.413
	- w/ SFT	0.0317	0.1880	0.0389	0.2366	24.77	2.144	6.62	3.745	27	4	85	0.511
	- w/ Planner only	0.0344	0.1992	0.0397	0.2557	32.57	2.624	6.12	5.321	30	13	105	0.423
	- w/ Realizer only	0.0354	0.2029	0.0400	0.2620	35.17	2.784	5.95	5.907	30	15	111	0.394
	- w/o diversity reward	0.0363	0.2066	0.0403	0.2684	37.77	2.943	5.79	6.526	29	18	118	0.365

Appendix

Table 23: Post-training objectives: economic performance and research diversity.

Market	Variant	Economic performance								Research diversity
Market	Variant	IC $\uparrow$	ICIR $\uparrow$	RIC $\uparrow$	RICIR $\uparrow$	ARR $\uparrow$	IR $\uparrow$	MDD $\downarrow$	CR $\uparrow$	M $\uparrow$	U $\uparrow$	C $\uparrow$	Pair corr. $\downarrow$
CSI300	- w/ improvement only	0.0281	0.1671	0.0371	0.2154	22.38	1.962	7.51	2.980	21	1	73	0.616
	- w/o balancing	0.0308	0.1814	0.0383	0.2296	23.95	2.084	6.96	3.441	25	3	81	0.546
	- w/o paired supervision	0.0273	0.1600	0.0365	0.2081	21.51	1.907	7.85	2.740	19	1	69	0.651
	full data SFT	0.0317	0.1880	0.0389	0.2366	24.77	2.144	6.62	3.745	27	4	85	0.511
CSI500	- w/ improvement only	0.0196	0.1195	0.0206	0.1350	-4.12	-0.312	13.56	-0.304	17	3	71	0.556

Appendix

Table 24: SFT-data ablations across four markets.

Method	Exec. $\uparrow$	Acc. $\uparrow$	Ret. $\uparrow$	Forms $\uparrow$	$\|\mathrm{RIC}\|\uparrow$
Untrained	151	112	17	146	0.0110
SFT	160	78	8	132	0.0108
GPT-5.5	160	104	35	144	0.0099
AlphaDiverse	160	118	56	150	0.0140

Appendix

Table 25: Factor screening on CSI300. Best and second-best values are bold and underlined, respectively.

Deployment	Input tokens	Output tokens	H200 GPU-hours
GPT-5.5	991,350	77,537	0.000
Grok-4.6	1,189,848	531,985	0.000
GLM-5.3	1,259,744	97,275	0.000
AlphaDiverse (local)	999,386	183,497	1.436

Appendix

Table 26: Recorded token usage and local GPU allocation per CSI300 research run.

Explore similar work

Aug 4, 2026cs.AI

AgonAlpha: Autonomous Alpha Discovery via Prompt Economy and Scalable Agentic Search

Language models can propose many plausible trading factors, but an autonomous research system must also allocate its evaluation budget, verify its own evidence, and preserve how each candidate was produced. We present AgonAlpha, an architecture that searches over frozen research artifacts---hypotheses, executable expressions, platform evidence, rationales, and review status---rather than formulas alone. To our knowledge, AgonAlpha is the first alpha-mining system to combine verified artifact search, a fresh-context adversarial reviewer with re-execution and veto authority, and pending-aware parallel budget allocation, together with a complete public evidence trail. Independent deployments on WorldQuant BRAIN produced SPECTACULAR-grade alphas across five users and six model backends, with Fitness reaching 9.50 and Sharpe reaching 3.48, while retaining prompt-to-expression provenance for every submission.

Weicheng Ye, Youran Sun, Xingyu Ren +3

Department of Physics, The Chinese University of Hong Kong, Hong Kong SAR, China · Department of Mathematics, University of Maryland, College Park, MD, USA · Department of Computer Science, University of Maryland, College Park, MD, USA

Algorithmic Trading Agentic Search Token Budget Allocation Data Mining

Ask PDF

Jul 29, 2026cs.AI

AlphaSchema: Exploring the Space of Trading Semantics for LLM-Based Alpha Mining

Automated alpha mining has increasingly adopted large language model (LLM) agents for factor generation and iterative discovery. However, existing LLM-based systems often delegate both factor construction and search decisions to the agent itself, without an explicit exploration space or a principled mechanism for navigating that space. As a result, exploration remains largely implicit and difficult to control or optimize systematically. We introduce AlphaSchema, which constructs and explores a structured space of trading semantics for alpha mining. Each point in this space is a schema plan composed of Event, Context, Qualities, Direction, and Output, specifying the semantics of a candidate factor before implementation. AlphaSchema decouples exploration from implementation: an LLM translates selected schema plans into executable factors, while evaluated rewards are accumulated to learn a surrogate model over the semantic space. An iterative selection mechanism uses this model to balance global exploration, surrogate-guided exploitation, and local mutation. Experiments on the Chinese stock market show that AlphaSchema discovers factor pools with strong predictive and portfolio performance. Further analyses show that the semantic search process navigates diverse regions while increasingly allocating evaluations toward high-reward regions, and that implementations of the same schema plans by different LLMs exhibit comparable predictive quality, suggesting that alpha mining quality is largely robust to the choice of LLM within our framework.

Jingyang Yi, Jian Yang, Yifei Jin +2

1X-Tech · 2Xtech-PandaAI-Waton Joint Lab · 3Monash University +2

Algorithmic Trading Schema

Ask PDF

Sep 8, 2026cs.AI

GoAnt: Quality-Diversity Multi-Agent Search for Alpha Factor Discovery in Market Microstructure Data

Automated alpha factor discovery searches symbolic trading signals from price-volume panels and order-book data under a fixed evaluation budget. Existing single- and multi-agent program-search systems can overfit predictive proxies that fail after execution costs and repeatedly explore redundant factor families, limiting execution robustness and behavioral diversity. We introduce GoAnt, a quality-diversity multi-agent search framework that combines non-communicating Explorer, Exploiter and Connector workers with a shared adaptive Mental Map and a compact Queen dispatcher. The Mental Map organizes candidates by leakage-free execution profiles and retains one elite per niche, while the Queen reallocates the evaluation budget from explicit search-state summaries. We also define a map-independent effective-yield protocol that counts high-quality, mutually nonredundant factors directly from each method's evaluation records, giving archive-based and map-free systems the same ruler. On real A-share microstructure data spanning 2023--2026, GoAnt reaches quality-weighted yields of 41.8 and 47.6 in price-volume and order-book settings, improving the strongest baseline by 57% and 97% under matched budgets. Its locked populations retain 0.64 and 0.67 of in-sample quality out of sample, compared with 0.61 and 0.63 for a static map.

Stella Zhao, Tommy Sha

University of Minnesota, Minneapolis, MN, USA · Stony Brook University, Stony Brook, NY, USA

Multi-Agent Planning Factor

Ask PDF