Research
Every research paper we have written up, newest paper first. New arrivals are screened automatically for topical fit with systematic trading; the ones that clear it are read, written up, and — where the idea is testable on our data — backtested. Particular focus on alpha hypotheses that can be formalised and tested. The Radar also covers portfolio construction, market risk and execution where the research is directly relevant to systematic investment processes. Follow new entries by RSS.
16,915 papers screened · 588 on the radar · 37 shown
LLM agents that act on live Web data must combine heterogeneous evidence, remember what they have done and improve from outcomes without weakening their own safeguards.
PAPER REPORTS · Live paper portfolio, Apr 8 to Sep 10, 2026 (102 trading days): - Return +5.0% vs SPY +12.4%. · Live paper portfolio, June 22 to July 29, 2026: -0.3%, vs SMH -24.6% over the same window.
Financial news can influence investment decisions long after publication. Timely news may reach an investor before prices fully adjust, whereas older news may already be reflected in prices.
PAPER REPORTS · LLM (GPT-4.1 mini) Market Data + Filtered News, 2023–2025 average of annual tests, net of 10 bp per dollar traded: -… · LLM Market Data + News + News Age, 2023–2025, net of 10 bp: 16.73% / 13.26% / 17.17% annualized at 0 / 1 / 4 week…
Formulaic alpha factor mining searches for symbolic expressions that transform market data into predictive signals. Recent LLM-based evolutionary factor mining methods use large language models as semantic variation operators.
PAPER REPORTS · CSI 500, test 2022-05-01 to 2026-04-30, top-20 combined signal, Qlib TopkDropout (top 50, drop 5, daily) net of open… · S&P 500, test 2022-05-01 to 2026-04-30, same backtest and cost settings: annualized return 27.00%, information ratio…
Which inputs drive a language model's financial forecast? We construct synthetic earnings announcements that independently reverse 20-day momentum, non-informational commentary tone, and whether fixed actual EPS beats a manipulated consensus.
Formulaic alpha mining seeks to discover executable expressions that capture predictive market signals in a vast, noisy search space.
PAPER REPORTS · AlphaRLM-8B CSI 300: annualized excess return 12.14% (0.1214±0.0011), MDD -6.15%, IC 0.0256, RankIC 0.0424, ICIR… · AlphaRLM-8B CSI 1000 (CSI 300-trained factor pool): annualized excess return 17.45%, MDD -11.98%, IC 0.0302, RankIC…
OUR BACKTEST · Sharpe 0.83 · Return +105.8% · Max DD -50.8%
The rapid growth of crypto markets is increasing investment opportunities, but also volatility risks. There is a need to automate portfolio construction to rapidly adapt investment decisions to rapidly changing markets.
PAPER REPORTS · Crew B (30-day rolling Sharpe max), out-of-sample per Table 6: expected return 10%, volatility 10%, Sharpe 0.72,… · Crew B (30-day window), out-of-sample per the window-sensitivity table: expected return -20.58%, volatility 39.88%,…
As option markets grow and AI advances, agentic systems for option trading are gaining increasing attention.
PAPER REPORTS · Total return 18.32% over 2025-03-03 to 2025-08-29 (six-month out-of-sample), 10 underlyings; fills at bid-ask midpoint… · Annualized Sharpe ratio 1.60 (same period and costs)
Financial LLM agents are often evaluated by comparing their end-to-end returns with those of a baseline and testing the paired difference against zero.
PAPER REPORTS · Full agent (A0) vs frozen-rule baseline, 723 earnings events, 2020-2024: gross deployment value +15.2 bps/event… · Same comparison net of a pre-stated 40 bps cost per nonzero event position: -1.2 bps/event (95% CI [-59.8, +54.3]),…
Large language models are moving from producing trading signals to writing the code that executes them.
In private wealth management, a manager delegating to artificial intelligence (AI) acts as the client's agent and the system's principal.
We study whether model diversity survives selection into trading. In synthetic markets with a fixed mixture of three language-model families, news presentation changes their representation among submitted orders.
Large language models are strong general-purpose code generators, but executable algorithmic trading remains a demanding specialization target: a model must translate a natural-language strategy specification into correct program logic for a specialized…
Self-evolving agents aim to turn research feedback into reusable skills, tools, and research rules. Whether these accumulated capabilities continue to improve later research requires controlled evaluation.
PAPER REPORTS · Mean final portfolio IC, development 2025-01-01 to 2025-12-31 (in-sample, per-bar cross-sectional Pearson vs next… · Mean final portfolio IC, OOS 2026-01-01 to 2026-06-30, with development-period weights frozen (predictive IC only; no…
Monetary-policy announcements and central-bank communications play a central role in foreign exchange markets, yet their qualitative, unstructured form makes their forecasting value difficult to quantify.
PAPER REPORTS · Directional hit rate, out-of-sample rolling windows, 2024-06-13 to 2026-06-12, no transaction costs considered: 66.08%… · Directional AUC 0.744 at 1h, 0.729 at 3h, 0.707 at 8h, 0.634 at 24h. Balanced accuracy 0.721, 0.698, 0.705, 0.626.
While symbolic regression (SR) has been successfully used in science to discover new equations, its use in financial valuation is hindered by several limitations.
PAPER REPORTS · Out-of-sample (last 20% of quarters) next-quarter price-level MAPE, not trading performance; no portfolio, Sharpe,…
OUR BACKTEST · Sharpe 0.41 · Return +37.0% · Max DD -44.3%
Language-model agents now run the whole of quantitative factor research: they propose investment factors, backtest them, select the survivors and retire them. We ask which of those jobs an agent should keep.
PAPER REPORTS · Certified (frozen-referee) book, 4-family library, 20 groups of 5 seeds, campaigns starting 2016-2019 through 2026-08:… · Deflated Sharpe ratio 0.49-0.69 against 108 groups.
Artificial intelligence (AI) now supports investment workflows from data and prediction through research, portfolios, execution, and tool use. Technical capability, however, is not evidence of investment profitability.
Large Language Models (LLMs) increasingly use user context such as memory, profiles, and role prompts to personalize their responses.
PAPER REPORTS · Q5-Q1 one-year (252 trading day) cumulative abnormal return relative to the S&P 500, sorting n~3,393 filings by… · Per-persona 12-model panel means: Value +5.63%, LongOnly +3.84%, ESG +3.44%, SellSide +3.30%, Credit +3.29%,…
Recent advances in LLM agents enable a new paradigm for asset pricing, which we call Agentic Empirical Asset Pricing (AEAP): systems that autonomously conduct the scientific discovery process itself. We define AEAP and identify its core building blocks.
PAPER REPORTS · SEADS mean per-factor OOS Sharpe 0.25 on Panel A (JKP) and 0.16 on Panel B (CRSP/Compustat), OOS windows 2020\u20132025… · SEADS productivity 14.0 (Panel A) / 13.8 (Panel B) admissions out of a 300-candidate budget
Quantitative trading is moving from isolated predictive models toward agentic workflows that combine reasoning, tool use, memory, and feedback.
Portfolio risk assessment ordinarily relies on reliable estimates of cross-asset return covariances, which are difficult to obtain in short, high-dimensional panels.
PAPER REPORTS · In-sample standardized variance percentile of the news-only allocation: 0.69%-1.33% across four prespecified capped… · Standardized in-sample variance 0.357, 8.3% below the equal-risk (inverse-volatility) benchmark and 35.6% above the…
OUR BACKTEST · Sharpe 0.47 · Return +44.2% · Max DD -29.9%
In financial markets, a sequential policy that reacts systematically to price movements may become predictable to other market participants.
PAPER REPORTS · Exposure-matched within-stock timing alpha, principal 10-minute Qwen3.5 price text: -45.7 bps per stock-day, 95% CI… · Chart-only 10m: -29.9 bps, CI [-33.7,-26.2], 14,937 stock-days; multimodal 10m: -48.9 bps, CI [-52.7,-45.0], 14,438…
Large language models (LLMs) are increasingly used to discover trading strategies, and much of the resulting literature shares a methodological weakness: many candidate strategies are generated, the best is reported, and neither look-ahead bias nor the…
PAPER REPORTS · Best gpt-4.1 discovery (E3, RSI x volume, 453-stock universe): design Sharpe 1.69 (2017-2021), evaluation Sharpe 0.18… · Best claude-sonnet-5 discovery: design Sharpe 0.44 (2017-2021), evaluation Sharpe -0.33 / -29% (2022-2025), DSR 0.18,…
OUR BACKTEST · Sharpe -0.10 · Return -10.8% · Max DD -47.8%
Large language models (LLMs) are increasingly used in investment decision-making, yet prior work shows that they exhibit systematic, model-specific investment preferences.
PAPER REPORTS · Weekly rebalanced equal-weight long-only top-100 portfolio from Qwen3-8B scores, 427 S&P 500 stocks, 29 signal weeks /… · Realized Sharpe of the top-100 portfolios is shown only graphically (Figure 5c, axis range roughly 2–4); no point…
OUR BACKTEST · Sharpe 0.52 · Return +57.3% · Max DD -40.6%
Costly LLM features matter only if calibration lets them affect the forecast. We document a failure of this link in a next-day risk study of two broad-market funds. Full-history scoring preceded the 2022 calibration.
PAPER REPORTS · Prespecified LLM importance feature: zero improvement, 95% interval [0,0], on all four endpoints (SPY/VIX binary and… · Signed LLM repair, SPY/VIX continuous variance: improvement -0.007452, 95% interval [-0.015888, -0.001066], Bonferroni…
Two old market sayings hold that news is already priced in by the time it is published, and that the rumor is bought while the news is sold. Both place the price move associated with a piece of news before and at publication rather than after it.
PAPER REPORTS · Fade small-cap launch/partnership news (short after positive, buy after negative; enter close of day +5, exit close of… · Short any covered small cap (sentiment-ignoring benchmark, 260,472 events, same windows, 2023-2026): 15.9% annualized,…
OUR BACKTEST · Sharpe -0.23 · Return -18.9% · Max DD -51.9%
This paper investigates whether textual tone derived from large language models (LLMs) can predict future stock returns. Using Korean news articles, we employ five LLMs to extract textual tones: BERT (KrFinBERT), DistilBERT, RoBERTa, ELECTRA, and Llama3.
PAPER REPORTS · BERT (KrFinBERT) equal-weighted quintile High-Low: 0.242% per day, t=3.772 (5.324% per month), May 2023-May 2024, gross… · BERT High-Low Fama-French 3-factor alpha 0.232% per day (t=3.582); 5-factor alpha 0.232% per day (t=3.566), same…
Large language models can extract richer signals from financial news than fixed sentiment lexicons, and recent work has explored feeding such signals into portfolio construction.
PAPER REPORTS · Sharpe 2.33, annualized return 95.9%, cumulative net 100.1%, max drawdown -18.3% — pure beta, GPT-4o mini, Student-t,… · Sharpe 2.44, annualized 100.0%, net 106.9%, max DD -18.3% — same configuration at 50 bps (2025)
We study how differences in AI-generated financial recommendations are transmitted into individual portfolio choices.
A new class of software systems is transforming investment analysis. Large language model agents assembled into collaborative team structures including analysts, researchers, and risk managers are increasingly deployed across financial markets.
PAPER REPORTS · Author's own prior human practice (not the AI framework): 127.17% return vs 50.67% benchmark within sixteen months,… · Ranked first among 276 comparable peer funds nationally through the 2020 market-stress period (peer average negative);…
This paper documents an applied natural-language-processing framework for measuring the tone of Brazilian Monetary Policy Committee (Copom) statements.
Agentic AI is gaining acceptance in asset management, but governance has not kept pace: 88% of surveyed finance professionals report no operational governance framework for agentic AI despite universal awareness of its deployment, and only 24 of 75 large U.S.
OUR BACKTEST · Sharpe -1.47 · Return -45.1% · Max DD -49.7%
We ask a representative sample to write prompts seeking spending and investing advice from LLMs, then simulate the lifetime effects of following the advice under realistic asset and labor market conditions.
I revisit the exchange rate disconnect puzzle, first documented by Meese and Rogoff (1983), using generative artificial intelligence (AI) to forecast currency returns based on economic fundamentals.
PAPER REPORTS · Annualized Sharpe ratio 0.594 at 48-month lookback (cross-sectional long top-2 / short bottom-2 of 9 currencies,… · Annualized Sharpe ratios 0.577 / 0.604 / 0.594 / 0.491 / 0.467 for 36 / 42 / 48 / 54 / 60-month lookbacks (2001-2024,…
Parent-order execution is a core problem in algorithmic trading, where the goal is to split a large order into smaller orders while reducing execution costs.
PAPER REPORTS · Aggressive setting, April 2026, DS-v4-f: value-weighted price performance -2.26 bps vs TWAP benchmark price, i.e. · Passive setting, April 2026, DS-v4-f: -3.92 bps wbp, +1.07 bps vs TWAP and +0.71 bps vs strongest baseline; 100%…
Recent advances in Generative AI have substantially improved financial sentiment analysis through post-trained financial large language models (LLMs).
PAPER REPORTS · FinSMART (static): cumulative return 264.9%, annualized return 91.5%, Sharpe 1.97, Sortino 2.40, Calmar 4.23, RankIC… · FinSMART (periodically retrained every 6 months): cumulative return 406.2%, annualized return 125.7%, Sharpe 2.41,…
OUR BACKTEST · Sharpe -0.46 · Return -48.2% · Max DD -75.2%
Powered by advances in LLMs and autonomous agents, deep research has become one of the most widely adopted agentic products. However, most deep research systems write general-purpose reports, which are inadequate for financial deep research.
PAPER REPORTS · FinanceHarness (Qwen3.6-27B backbone): overall rubric score 32.4%, pre-cutoff 45.7%, post-cutoff 11.8%, bootstrap SE… · FinanceHarness + GRPO (RFT): overall 32.8%, pre-cutoff 46.2%, post-cutoff 12.1% (same backbone and benchmark)