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 · 99 shown
Portfolio optimization techniques often use historical data to estimate parameters for Markowitz-type or ML models. Effective allocation requires a way to align the available data with conditions expected over the investment horizon.
PAPER REPORTS · Markowitz, lambda=0 (historical mean only, no future information in the view): net Sharpe 0.019 (gross 0.050), net MDD… · LSTM mean of 100 seeds, lambda=0: net Sharpe -0.021 (gross 0.037), net MDD -24.7%, final wealth 0.978.
Optimal capital allocation in financial markets has long been a challenging task due to the dynamic nature of market conditions, risk factors, and the complexity of financial data.
PAPER REPORTS · Annualized Sharpe ratio 1.91, computed from daily portfolio returns with 252 trading days and a 2% annual risk-free… · Cumulative return 1.12. The test period is not stated beyond the 250 trading days in Fig 7, and it is unclear whether…
Autoregressive foundation models offer state-of-the-art capabilities to simulate the market by generating the message flow of limit order books (LOBs).
The rapid advancement of generative artificial intelligence has catalyzed a paradigm shift in quantitative finance, transitioning from traditional point forecasting toward comprehensive financial time-series generation (FTSG) across extrapolation, imputation,…
Estimating extreme risk in reinforcement learning becomes increasingly difficult as the confidence level increases, since only a small number of observations contribute to empirical estimates of Conditional Value-at-Risk (CVaR).
PAPER REPORTS · Simulated Merton jump-diffusion, K=140 call, 63 daily steps; transaction costs not mentioned. · Same setting, EVT-trained policy hedging error on 3,000 paths conditional on unhedged payoff > 0.999 quantile: mean…
Realistic market impact is essential for generative limit order book (LOB) models to serve as world models for trading agents. If a simulator misprices this mechanism, agents learn policies that are optimal in the simulator but wrong in the real market.
Language models that read financial news are usually trained on text from the same years used to test them, so a backtest can reward hindsight.
PAPER REPORTS · Monthly, price+text combination across all readers and models, 63-day shocks, test 2022–2024, net of 10bp per dollar… · Same monthly buffered portfolio: net return 12.4%/yr, volatility 10.4%, max drawdown 8%, turnover 60% per month,…
OUR BACKTEST · Sharpe -0.73 · Return -19.9% · Max DD -34.5%
Automated markets increasingly host autonomous, learning trading agents and are powered by formulaic automated market makers (AMMs).
PAPER REPORTS · Rolling 7-day annualized Sharpe ratio of net strategy returns ranges roughly between -20 and +20 for every agent…
We study the conditional distribution of limit-order volume submitted over the next five minutes, given the current order book and recent order flow. The task requires capturing price placement, total volume, and predictive uncertainty jointly.
Score-based diffusion methods usually perturb data with Brownian Gaussian noise, corresponding to a light-tailed forward stochastic differential equation (SDE).
PAPER REPORTS · Feishu A-share OOS (one-year out-of-sample period, dates not stated), StochLOB (jump-Poisson) constrained long-only… · BTC chronological held-out directional classification (test 2026-07-10 to 2026-07-13, no trading costs): StochLOB…
Stock price forecasting is considered difficult due to the volatility of financial markets. A reliable stock price forecasting model is critical for making sound investing decisions while managing financial risks and portfolios.
Physics-informed neural networks provide a flexible framework for learning parametric PDE solution families without labeled data, yet preserving accurate derivatives across heterogeneous regimes remains challenging.
A conditional normalizing flow can produce extreme scenarios even when its point forecasts appear reasonable. We compare two fitted models for daily Treasury yield changes at eleven maturities reconstructed from Federal Reserve curve parameters.
PAPER REPORTS · Stabilized flow, development windows (5×60 dates, 2002–2026; forecast metrics only, no trading strategy or costs):… · Stabilized flow, subsequent window April 13–July 8, 2026 (frozen checkpoints): scenario-mean MSE 0.3351, 90% coverage…
OUR BACKTEST · Sharpe 0.03 · Return +0.2% · Max DD -5.6%
Market making under realistic order-flow dynamics is challenging due to the clustered, self-exciting nature of order arrivals, which are naturally modeled by non-Markovian Hawkes processes.
PAPER REPORTS · HSN mean cumulative episodic reward, simulated (no market data), multi-asset cross-excitation cs=0.7: 0.0548±.0007 over… · HSN, simulated power-law kernel ε=0.15: 0.0216±.0004 mean cumulative reward, +32.5% vs the best baseline (p<0.001).
A path indistinguishable from its own reverse carries no arrow of time, so time irreversibility is a necessary condition for a directed mechanism to be readable from trajectories alone.
Temporal point processes are used both to measure market mechanisms, for example how fast prices absorb news and how much of the reaction is the market responding to itself, and to predict the next event.
We examine climate variables in a pipeline for dynamic Value-at-Risk (VaR) of the SRI-KEHATI ESG index and the IHSG (Jakarta Composite Index).
PAPER REPORTS · Main QRNN backtest, test N=300 (Nov 2023-Dec 2025), VaR violation rates and Kupiec p: SRI-KEHATI tau=0.01 0.33% (1… · Main QRNN DM statistics on pinball loss (positive = QRNN better), all p<0.001. Vs HS-VaR: SRK 4.31 / 6.43 / 7.23 and…
How can we identify stable patterns to accurately predict stock price movements amidst severe market noise? Predicting stock price movements is a crucial problem in financial data mining, and has attracted significant attention from researchers and financial…
PAPER REPORTS · COVID-19 crisis (test 2020-01 to 2020-03), threshold long-short, 6 bps cost on sells, zero risk-free rate: S&P500 ASR… · AI rally (test 2024-01 to 2024-06), same protocol: S&P500 ASR 2.842, RMDD 6.8%; CSI300 (long-only) ASR 1.653, RMDD…
Correlation matrices combine nonlinear positive-definiteness constraints with a simultaneous row-column permutation symmetry. Their generation therefore involves two distinct problems.
OUR BACKTEST · Sharpe 0.00 · Return +0.0% · Max DD 0.0%
This paper studies exit timing for SPX short-straddle portfolios using a two-stage supervised and reinforcement-learning framework.
PAPER REPORTS · Supervised MLP exit policy, OOS 2024-2026 (T≈2.42y), stress-budget compounding sizing, bid entry/ask exit plus… · AC+MLP, 7 seeds, OOS 2024-2026, same costs: cumulative return 37.54% ± 0.92%, ARC 14.05% ± 0.32%, ASD 7.97% ± 0.14%,…
Abstract Time series forecasting models play a crucial role as decision support tools in various real-world domains.
PAPER REPORTS · Multi-stock swing trading (17 stocks, equal budget, 3-day confirmation), uptrend Jul 1–Dec 31 2021: average gain 7.64%… · Multi-stock swing trading, downtrend Jan 1–Sep 20 2022: average gain -0.34% across Max Shares 1/5/10 (-7.16%, 3.61%,…
When a trader executes a large order, the relevant question is not \emph{what will the price be?} but \emph{what would the price have been had I not traded?} Market impact is thus a causal quantity, yet classical models (Almgren--Chriss, propagators) regress…
Abstract This study examines the predictive relationship between sentiment extracted from Google News headlines and the log spreads of cointegrated or highly correlated stock pairs, aiming to enhance pairs trading profitability.
PAPER REPORTS · MA_V, Full-Predicted gn_a: return 301.52%, Sortino 1.029, Sharpe 0.693, Max DD 0.016, 181 trades (68-day test period,… · MA_V, Full-Predicted rand_a (random sentiment): return 314.36%, Sortino 0.446, Sharpe 0.686, Max DD 0.021, 325 trades…
Wavelet-based financial forecasters typically use the transform only to denoise, or reduce it to a single spectral snapshot at the forecast origin, and the convolution that produces the coefficients is usually bilateral, so it can read past the forecast…
PAPER REPORTS · S&P 500 DSTNet asymmetric risk-off allocation, test 2023-06-01 to 2024-05-31, 10 bps per unit turnover, idealised… · NYSE DSTNet risk-off, same period, 10 bps: Sharpe 1.80, MaxDD 5.65%, CumRet 11.31%, exposure 56.77%, total turnover…
OUR BACKTEST · Sharpe 0.46 · Return +21.1% · Max DD -24.7%
Shallow models are special cases of deep models, and deep models theoretically have the potential to outperform the shallow ones. However, the existing empirical asset pricing literature provides strong benchmarks for shallow models.
PAPER REPORTS · Deep ResNet+ (depths 11-20, spec-averaged): annualized VW decile long-short SR 2.07, monthly mean 3.48%, SD 5.82%, Feb… · Deep ResNet+ FF5 alpha 3.35% per month (significant at 1%, Newey-West 12 lags), Feb 1987-Dec 2023, gross of costs
Learning effective representations helps characterize the structure and dynamics of equity markets from financial data with a low signal-to-noise ratio.
PAPER REPORTS · CN_ALL top-30 equal-weighted long-only, 21-day overlapping batches, test Jan 2024–June 2026, net of 5 bps per unit… · US_ALL top-30 minus bottom-30 long-short, same horizon and costs, test Jan 2024–June 2026: Net annualized return…
We develop a posteriori primal-dual bounds for numerical approximations of European option prices in the Uncertain Volatility Model.
Deep hedging learns trading policies from historical or simulated market trajectories, yet under nonstationarity these training paths may not represent future market conditions.
PAPER REPORTS · Nonstationary Heston (simulated), European call, CVaR_0.95 of hedging loss, no transaction costs: WRAP-W far-future… · Nonstationary Heston (simulated) with proportional costs c=0.005: WRAP-U far-future CVaR 8.349±0.189 vs ERM-U…
Formulaic alpha discovery seeks symbolic expressions that predict cross-sectional asset returns.
PAPER REPORTS · CSI300, test 2023-2025, long-only top-20% daily cohorts held 20 days, net of 5bp buy / 15bp sell, mean of 5 runs: IC… · CSI500, test 2023-2025, same protocol and costs: IC 0.0444, ICIR 0.2855, RankIC 0.0763, RankICIR 0.4578, annualized…
This paper develops a reproducible framework for predicting gold-price direction and converting the forecasts into a currency-hedged SHFE-COMEX spread-trading strategy.
PAPER REPORTS · Hedged (USD/CNH) walk-forward strategy: annual return 14.8%, volatility 7.1%, Sharpe 2.08, max drawdown 4.2%, win rate… · Unhedged version, same setup: annual return 11.2%, volatility 9.0%, Sharpe 1.24, max drawdown 12.7%, win rate 61.3%, 52…
networks and Transformer models for financial volatility forecasting, with specific applications to the CBOE Volatility Index (VIX) and cryptocurrency markets (Bitcoin, Ethereum).
The complex and nonlinear behaviors of financial markets make it exceedingly challenging for average investors to evaluate the performance of a specific asset or make meaningful predictions about its future behavior.
PAPER REPORTS · Directional hit rate (MDA) of the LSTM forecaster on a random 80/20 holdout validation set (27,433 samples, 159 stocks,… · Diebold–Mariano statistics vs baselines on the same validation set: 3.42 (p = 0.001) vs persistence and 2.15 (p =…
Deep reinforcement-learning policies for order execution can vary substantially across training seeds, so apparent architectural gains may reflect favourable training realisations rather than reproducible properties of the architecture.
PAPER REPORTS · Mean implementation shortfall, out-of-sample Jan 16–20 2023, frictionless replay with no commissions or market impact,… · Per-run CVaR-95 under the same setup: DDQL 110.8, MoE K=8 122.2, DDQL-cap8 104.1 bps.
Neural network-based approaches have emerged as efficient alternatives to traditional optimization-based procedures for the calibration of stochastic volatility models.
Short-horizon price-trend prediction from limit order books in equity and intraday electricity markets requires models that combine predictive quality with low single-sample latency and a small serialized model size to keep pace with rapid and continuous…
PAPER REPORTS · Classification performance only. These are not trading returns: no Sharpe, PnL or cost assumption is reported, and the… · MBOFormer test macro-F1 (mean of 3 seeds), h=10/20/50/100: - EPEX (3 test days, March 2021): 0.651/0.577/0.517/0.472 -…
Abstract Estimating value-at-risk on time series data with potentially heteroskedastic dynamics is a highly challenging task.
PAPER REPORTS · LSTM VaR estimator, Fama-French 25 BM/OP portfolios, VaR 1%, n=250, test period ending 08.03.2019 (~500 backtest days):… · LSTM VaR estimator, Fama-French 25 portfolios, VaR 1%, n=250, test period ending 28.02.2022 incl.
We develop a spouse-protected tontine in which a first death changes the household state but generates no pool transfer.
PAPER REPORTS · Out-of-sample learned policy, payment-seeking (λ=0.3): expected cumulative real payments 1860.4 (thousand AUD over 35… · Out-of-sample learned policy, intermediate (λ=1): expected cumulative real payments 1704.8, CVaR_0.95 terminal…
Machine-learning signals built from financial text treat what institutions say, and what the media repeat, as evidence about value. But whoever shapes a narrative may be trading against it.
PAPER REPORTS · Linear model, raw voices + theory features: out-of-sample cross-sectional IC 0.171±0.018 (simulated main market, 5… · Gradient boosting, raw voices + theory features: IC 0.118±0.019 (simulated main market, 5 seeds; costs not stated)
Power systems with increasing variable renewable generation face greater uncertainty in scheduling and balancing. Intraday and balancing electricity markets facilitate position adjustments and real-time balancing close to delivery.
In this work, we study portfolio optimization under the stochastic discount factor (SDF) framework by learning market state representations that capture the underlying risk structures of financial data.
PAPER REPORTS · Sharpe ratio 1.062 (out-of-sample rolling 1-year test windows within June 2011–June 2020, 7-day rebalancing;… · Sortino ratio 1.352; Calmar ratio 0.924 (same setting; costs not stated)
We introduce a deep kernel hedging framework that combines the flexibility of deep learning with the structural inductive bias of kernel methods.
PAPER REPORTS · S&P500 out-of-sample (Jan 1 2023 to May 31 2026), ATM European call, quadratic hedging with order-3 signature features,… · S&P500 out-of-sample (Jan 1 2023 to May 31 2026), ATM European call, CVaR hedging, frictionless.
OUR BACKTEST · Sharpe 2.30 · Return +22.0% · Max DD -2.8%
Predicting stock price movements is highly challenging due to the noisy, nonlinear, and multi-factor nature of financial markets.
OUR BACKTEST · Sharpe 0.70 · Return +20.9% · Max DD -6.5%
News may reveal systematic risk, but whether its context enhances the construction of systematic risk factors is still unclear.
PAPER REPORTS · Spherical ST + sparse-soft c-TF-IDF + multi-horizon IPCA, L=60: annualized OOS excess-return Sharpe 1.0334, Sep 2020 to… · Same raw spherical portfolio: annualized volatility 81.68%, maximum drawdown 62.23%; not risk-managed, not…
We present an end-to-end deep learning framework for systematic options trading that directly embeds hedging behavior through explicit control of portfolio-level risk exposures.
PAPER REPORTS · DP (L1), α=10^2: annualized Sharpe 2.339, E[Return] 0.436, Vol 0.186, MDD 0.210, Sortino 3.583, Calmar 2.078, hit rate… · ENP (L2), α=10^1: Sharpe 2.260, E[Return] 0.409, MDD 0.236, hit rate 0.638. Zero costs.
We study simple predictable processes whose coefficients are represented by neural networks. On finite measure spaces, we establish density results for neural networks in Orlicz spaces.
Financial time series evolve across multiple temporal resolutions, challenging forecasting systems to incorporate newly available information without repeatedly recomputing unchanged representations.
PAPER REPORTS · HARN test-split forecast error (not a trading strategy; no costs or P&L), AAPL M15, 2025-03-06 to 2026-08-10: MAE… · HARN test-split forecast error, EURUSD M15, 2025-12-19 to 2026-07-29: MAE 0.000306±0.000002, RMSE 0.000467, sMAPE…
Stock price forecasting is a long-standing challenge in computational finance, driven by the inherent randomness of markets and complex temporal patterns.
PAPER REPORTS · CSI300 test (2017-01-01 to 2020-08-01), Qlib top-50 long, 20 replacements/day: excess annualized return 0.1809, excess… · CSI500 test (2017-01-01 to 2020-08-01), same strategy: excess annualized return 0.1529, excess information ratio…
Financial language models can transform unstructured firm-specific news into structured decision signals, but financial AI research lacks an integrated deployment framework for evaluating whether those signals remain useful in financial decision systems.
PAPER REPORTS · Long-short net Sharpe 2.85 (LLaMA-3, value-weighted quintile, net of 5bp one-way costs, 10% ADV cap, Jun 2024-Jan 2026) · Mean daily net return 0.34%, daily vol 1.89%, max drawdown -12.3%, cumulative long-short return 180% (LLaMA-3, same…
We study optimal stopping under dynamic risk measures with simultaneous ambiguity in the probability model and the discount rate.
Modeling the dynamics of option implied volatility surface (IVS) is crucial for pricing, hedging, and risk-managing option portfolios.