financial data engineers

Abstract: Financial instruction-data pipelines must distinguish content validity from permission to release a particular dataset. Arya Forge combines typed records, current-content checks, and approval coverage bound to a dataset manifest. We evaluate its release boundaries in a financial-data systems case study. An integrity audit covers 100,000 synthetic instruction records; a separate arithmetic implementation agrees with all 15,000 numeric records across twelve formula families. In 240 scenarios with 2,400 treatment rows, adding financial and release-record checks to a short-instruction filtering recipe rejects all injected numeric faults and policy holds while retaining 220 of 240 clean controls. The twenty lost controls comprise every tested cash-conversion-cycle example. Nineteen persisted-release cases show that eleven selected negative cases are blocked, four equivalent representations are rejected, and changes to external evidence after binding are outside ordinary Gold promotion's checks. Local processing measurements and controlled checkpoint resumptions characterize cost and recovery through 100,000 input rows. The results expose a tradeoff between strict producer-contract enforcement and useful-representation retention, and identify the protected inputs required for manifest-bound approval. These findings apply to the tested implementation and workloads; they do not establish general financial correctness, authenticated approval, expert-validated quality, or downstream model improvement.

Abstract: Reliable evaluation of LLM-based alpha mining requires more than positive backtest summaries. We audit a preserved project at three levels: historical run counters, outcome retention, and a separately scoped public reanalysis. Three historical runs report 67, 20, and 11 keep decisions, but the comparison mixes metric populations: a mean over kept candidates is compared with a mean of per-round maxima, and successful evaluations with all returned outcomes. These artifacts cannot identify an architecture advantage. A released 1,837-expression catalog also excludes non-positive OOS outcomes, making its zero-reversal rate non-informative. We then account for all 1,047 unique expressions in a frozen, IS-screened preserved pool using forty A-shares and 2024-2025 rank associations. A versioned public-operator correction yields 923 scorable records, 99 with insufficient valid dates, and 25 unsupported, without modifying the other 990 prior diagnostics. Among scorable records, 671 have positive and 252 negative signed mean rank-IC. Simple untrained references also show positive associations, so this distribution does not establish incremental agent skill or profitability. The case study motivates explicit metric grains, immutable cohort accounting, and provenance-aware interpretation rather than architectural rankings from incomplete records.

Abstract: A market maker provides liquidity by quoting prices to buy and sell an asset, earning the spread while managing inventory risk. We study market makers trained with multi-agent reinforcement learning on prediction markets, specifically Polymarket, whose thin order books pose distinct challenges. Three challenges underpin this research: adverse selection may cause a policy trained on real data to stop quoting; the consequences of a quote can arrive many steps after the action; and prediction-market L2 data does not identify individual orders needed for simulation. We address these challenges by reconstructing order-level LOBSTER-format data from Polymarket and training agents in JaxMARL-HFT on 417 real markets. We enrich the baseline market maker's observation space with the book and recent flow, redesign its actions and inventory controls, and propose a cooperative multi-agent decomposition into a quoter and a risk manager trained using recurrent Independent Proximal Policy Optimisation (IPPO) with RNN policies. We investigate reward shaping, particularly reward blending and post-fill markout, to mitigate credit-assignment difficulties between the agents. Our redesigned single-agent market maker quotes more frequently and incurs smaller held-out losses than the retrained JaxMARL-HFT baseline, although the contributions of jointly changed observations, actions and risk constraints are not separately identified. Decomposition alone performs poorly; careful reward design substantially improves the desk relative to the unshaped split and yields our best result. Models are evaluated on realised end-of-episode portfolio value on held-out markets. None establishes positive average portfolio value in periodic held-out evaluation, and each configuration was trained with one seed. The environment, data and protocol lay a foundation for makers that quote many linked markets at once.

Abstract: Sector rotation is a common decision in benchmark-relative portfolio management, where portfolio managers rank sectors for overweight and underweight positions. We formulate this task as ranking sector actions by expected future risk-adjusted active performance, measured with forward sector-benchmark Sharpe ratios. Such rankings should preserve temporal information, express model outputs through economically familiar signal characteristics, account for changes in signal relevance across macroeconomic environments, and capture dependencies among sectors. We propose RegimeRank, an interpretable sector-ranking model that represents fundamental and technical signals through expert-defined temporal characteristics, conditions them on learned macro factors via cross-attention, and models cross-sector dependencies through self-attention. Expected-gradient attributions identify the contributions of individual signal-transformation pairs, while macroeconomic and cross-sector attention maps provide diagnostics of the learned conditioning and relational structure. We evaluate RegimeRank on weekly S&P 500 data from 2002 to 2024 under combinatorial purged cross-validation, where it outperforms representative ranking and machine-learning baselines across multiple ranking metrics. Ablation studies support the contributions of latent macro states, regime conditioning, and cross-sector modelling, while attribution and attention analyses illustrate the temporal characteristics, macro states, and sector relationships underlying its rankings.

Abstract: Stock price forecasting is hindered by the heteroscedastic nature of financial time series, in which turbulent and tranquil regimes alternate. Although the Long Short-Term Memory (LSTM) network is widely used for this task, its gates are oblivious to such volatility variation, so memory retention remains fixed regardless of market conditions. A common remedy is to supply volatility estimates as an additional input feature, yet this leaves the internal memory dynamics untouched. Instead, volatility can serve as a direct regulatory signal over the gate activation itself. This paper presents the Gsigmoid-LSTM model, which embeds Generalized Autoregressive Conditional Heteroscedasticity (GARCH(1,1))-derived volatility estimates into the LSTM forgetting gate through the Gsigmoid activation function, a volatility-adaptive variant of the sigmoid that combines a volatility adjustment factor with slope regulation via a learnable parameter. The model thereby automatically prioritizes recent information during turbulent periods while preserving long-term dependencies in stable markets. In the engineering application of daily stock price forecasting, extensive experiments, including activation function comparisons, baseline evaluations, and ablation studies, demonstrate that this structural gating modification effectively improves forecasting accuracy and robustness. The proposed model significantly outperforms a feature-level fusion baseline carrying the same volatility signal on all five datasets, as confirmed by Diebold-Mariano tests with all p<0.01, attains the best average rank of 1.00 in the Friedman test across nine models, and achieves the lowest MSE and MAE on all five datasets.

Abstract: 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. We build news topics from NoLBERT, an encoder pretrained only on text published between 1976 and 1995, adapted into a sentence encoder by contrastive fine-tuning on passages of the same article, with no labels or returns. We compare it with LDA, MiniLM and FinBERT readers on 1.26 million articles from 2012 to 2024, convert each reader's topics into attention shocks and stock exposures, and forecast the returns of the 1,000 largest US stocks with eight models at three rebalancing frequencies. Models are fitted on 2013-2019, choices are made on 2020-2021, and 2022-2024 is held out. The experiment with the best validation score loses money at every frequency. Measuring attention shocks against the past 63 trading days instead of 5 raises the test rank IC of every reader. NoLBERT gives the most stable topics and forecasts about as well as LDA. An equal-weight combination of all readers and models earns a monthly net Sharpe ratio of 1.12, significant against zero before adjusting for the search but not against a price-only benchmark.

Abstract: Score-based diffusion methods usually perturb data with Brownian Gaussian noise, corresponding to a light-tailed forward stochastic differential equation (SDE). For market time series this SDE can be a poor abstraction: limit order book (LOB) order flow is heavy-tailed, bursty, and occasionally jump-like. We study STOCHLOB, a training-only score-matching framework that changes the forward diffusion corruption law while leaving the deployed predictor deterministic. Instead of Brownian-only Gaussian corruption, the noising process uses two non-Gaussian SDE-inspired alternatives: jump-Lévy noise through a compound-Poisson variance mixture, and clipped $α$-stable noise through a positive-stable Gaussian scale mixture. These corruptions inject finite jumps and heavy-tailed residuals into training without requiring reverse-time diffusion sampling at inference. The resulting score target has a useful interpretation: its radial factor $h_t(r)$ is the posterior expected corruption precision, so large residuals are weighted according to the latent corruption scale that most plausibly generated them. On chronologically held-out Coinbase LOB data, the proposed variants remain competitive on uncorrupted BTC and altcoin test horizons and improve several realized tail-window stress diagnostics. An exploratory Feishu A-share transfer study further tests whether the probability outputs remain usable under transaction costs and trading-rule constraints. Code and configuration files are available at https://github.com/ArshiAbolghasemi/Penny.

Abstract: Autoregressive transformers trained on limit order book (LOB) message streams are increasingly proposed as foundation models for market microstructure. Because the generative process of real markets is unknown, these models can be judged only by the realism of their output, leaving open what they actually learn. We make the question answerable by inverting the setup: we train GPT-style models on order flow from simulators whose latent parameters are known, so that the Bayes-optimal predictor, the generating state, and the exact data entropy serve as references. Measured against them, next-token pretraining on order flow implements amortized Bayesian inference over market regimes. The learned predictive matches the exact Bayes optimum to within a millinat per token and approaches the oracle floor even on a nonstationary family whose latent state drifts within an episode, a regime that requires filtering. The latent regime is linearly decodable from frozen activations, well above baselines, so a frozen-model read-out calibrates simulator parameters more accurately than windowed maximum likelihood at a fraction of the cost. In our experiments these capabilities are not produced by scale: capacity and token volume saturate early, and it is the diversity of training regimes, not the token count, that determines generalization to unseen markets. On real NASDAQ data, zero-shot transfer is set by how well the simulator prior covers real flow: an enriched mixture nearly closes the gap to an in-domain baseline, and pretraining beats from-scratch training at every budget across held-out tickers. Known ground truth turns average next-token loss, which is uninformative near the entropy floor, into an exact comparison against the Bayes-optimal and oracle predictors.

Abstract: Comparing financial disclosures across languages requires identifying shared accounting concepts and recovering all numerical facts that express them. Existing benchmarks evaluate aspects of these capabilities, but provide limited insight into whether both succeed on the same multilingual evidence. In particular, it remains unclear whether language models can recognize these concepts across languages and exhaustively recover their corresponding numerical evidence from the same contexts. We introduce MULTIFINGROUND, a benchmark designed to investigate this recognition-grounding gap in multilingual financial reports. Constructed from FY2025 disclosures of 90 issuers in English, French, and Greek, MULTIFINGROUND aligns independently authored reports through a shared accounting taxonomy and provides traceable fact-level annotations. The benchmark evaluates financial fact identification, concept recognition, exhaustive fact selection, and concept discrimination without candidate labels. Across 15 evaluated models, eight API and open-weight reference models achieve 97.3-99.4\% concept recognition accuracy, yet miss at least one target fact in 26.6-41.9\% of correctly recognized groups. Further analyses reveal substantial variation in concept discrimination and limited gains from the evaluated finance-specialized models. Our findings demonstrate that strong concept recognition can coexist with incomplete evidence recovery, highlighting the importance of evaluating semantic recognition and evidence completeness together.

Abstract: Automated alpha factor discovery is critical for modern quantitative research. While large language models (LLMs) have increasingly been integrated into quantitative workflows, existing agent-based approaches remain highly susceptible to severe backtest overfitting. To address this challenge, we propose EvoQuantLab, a training-free, dual-loop self-evolving framework for robust factor discovery. Instead of relying entirely on LLM generation, EvoQuantLab deploys an inner genetic-programming (GP) loop that dramatically improves token efficiency while evolving typed factor expressions under structural and diversity constraints. An outer diagnosis loop then systematically audits evidence across cycles to drive protocol-level improvements. Furthermore, our evaluation pipeline applies multi-metric validation to rigorously filter out spurious factors before deployment. Extensive out-of-sample experiments across both daily and hourly frequencies -- covering equities and cryptocurrency -- demonstrate the robustness of EvoQuantLab. Specifically, on daily CSI 500 and S&P 500 tasks, EvoQuantLab achieves leading annualized returns of 18.34% and 21.30%, consistently outperforming existing agent baselines. On high-turnover hourly cryptocurrency data, EvoQuantLab delivers a 37.91% annualized return while effectively avoiding execution collapse.

Abstract: Synthetic tabular generators are commonly optimized for distributional fidelity, but statistical similarity alone does not guarantee preservation of causal effects. In this paper, we study whether causal fidelity can be improved directly within a fully generative tabular model. Causal Fidelity is defined with respect to a target estimand as the discrepancy between inferential distributions obtained from real and synthetic data, and theoretical results show that high statistical fidelity does not generally imply high causal fidelity. We then propose a causal-fidelity-aware training framework which adds a causal discrepancy penalty to the generative objective. The framework is instantiated with a causal-penalized TabDDPM and optimized using an on-policy score-function estimator. We further establish conditions under which causal regularization improves expected causal fidelity. Experiments across diverse treatment-effect simulations and two benchmark datasets evaluate the ability of our method to improve causal fidelity while preserving competitive statistical fidelity.

Abstract: Risk regulation imposes directional constraints on scores; we adopt their strict per-input form -- the score monotone non-decreasing in every exposure input -- as a normative commitment. Deployed pipelines -- monotone hand-crafted aggregates feeding sign-constrained gradient boosting -- already satisfy it by composition, so constrained-versus-unconstrained comparisons price a guarantee the incumbent has for free. We instead hold admissibility fixed on both sides and measure what learning the aggregation is worth. Our instrument is a recurrent network whose state is classical risk statistics (an exponentially weighted moving average and a high-water mark with learned transforms), monotone by construction in every input and per MC-dropout sample. The central finding, by functional regression, is a subsumption boundary: a learned monotone channel reproduces the geometrically weighted separable family of hand-crafted statistics, one channel per member, to Spearman $ρ\ge 0.996$, approximates window statistics with measurable ceilings, and fails at consecutivity ($ρ= 0.924$) and time localization (0.628), both structural, and at the exposure floor (0.829), a learnability boundary. One explicit admissible basis repairs each failure (rank correlation 1.000). In or near the separable family, learned and engineered aggregation are substitutes, and the learned channel is never statistically behind at full sample size and specified capacity. Its advantages are incumbent-specific: a committed grid pays up to 0.019 AUC in decay regions it leaves uncovered (the learned channel stays within 0.004 of the strongest engineered consumer at every swept point); the highest-dimensional comparator degrades fastest with scarce data; and beyond the training support, grid-fed tree-ensemble scores go flat while a strictly increasing head keeps ranking. No single incumbent is dominated on all three axes.

Abstract: Account takeover (ATO) fraud is a growing threat to digital banking, requiring effective detection while minimizing friction for legitimate customers. Production systems predominantly rely on tabular models that score sessions in isolation, discarding the relational structure of the underlying interaction network. Although graph-based models exploit relationships among sessions and network entities, they primarily reason over local neighborhoods and therefore capture only part of the problem: fraud risk depends jointly on the local relational structure surrounding a session and the evolving global state of the fraud ecosystem. We present HERMES (HEterogeneous Relational Micro--macro graph transformer Encoder for high-risk Sessions), a dual-scale architecture that jointly models these complementary scales of information. Micro-GT captures local heterogeneous graph structure through structured, relation-aware attention over a temporally safe session neighborhood. Complementing this local representation, Macro-GT models ecosystem-level context using non-anticipative climate tokens that summarize fraud dynamics, platform shifts, and infrastructure reuse, together with adaptive class prototypes that track representative fraud and benign session patterns over time. Evaluated on more than 130 million high-risk transaction sessions from a leading U.S. financial institution, HERMES consistently outperforms production and strong graph-based baselines, achieving a 44.44% relative reduction in customer friction and a 24.66% relative improvement in fraud recall over the production system. Ablation and temporal-stability analyses further demonstrate complementary gains from local relational modeling and global ecosystem context across changing fraud regimes.

Abstract: General-purpose Text-to-SQL systems achieve strong performance on academic benchmarks like Spider and BIRD, where schemas are relatively shallow and column values are often human readable. In production financial databases, where concepts are stored as opaque integer keys rather than human-readable strings, these methods fall below 50%, as even simple queries require multiple joins and filter predicates reference opaque IDs. We present Financial LINking Text-to-SQL (FLINT), a domain-specialized Text-to-SQL system that closes this gap through three key components: (1) a lookup agent that dynamically resolves natural-language concepts to question-specific reference table constraints, (2) embedding-based retrieval of structurally similar query templates from a compact, expert-authored bank, and (3) schema linking that prunes a large table schema to the relevant subset by traversing foreign-key chains, rather than relying on name similarity alone. We evaluate on two datasets totaling 359 questions over production financial schemas. FLINT outperforms various state-of-the-art baselines using the same LLM. The system is deployed in production as part of a financial data retrieval service.

Abstract: We test the stability, reproducibility and investment outcomes of language-model stock rankings. Four models and five numerical comparators share a portfolio engine over 72 monthly holding periods in the Shanghai Stock Exchange (SSE) 50, China Securities Index (CSI) 300 and CSI 500. Rankings use nine characteristics, and five repeated SSE 50 runs measure variation under identical inputs. Linear rules fitted to development-period model preferences are frozen before unseen-month, larger-pool and controlled-intervention tests. Their mean Spearman agreement with model rankings is 0.923-0.984 in the SSE 50 and 0.795-0.985 after transfer. Aggregate rank-change error falls relative to a zero-change prediction in twelve archived feature-group comparisons and eight matched single-feature comparisons, with Holm adjustments applied in separate nine-plus-three and six-plus-two families. Prediction of individual entries and exits remains weak (event Jaccard 0.000-0.125). Historical mean model compound annual growth rates range from 6.26% to 11.17%. At 10 basis points per side and six-month blocks, the twelve-comparison model-minus-rule return family and factor-controlled associations yield no adjusted finding. Three higher-cost, twelve-month-block comparisons favor a Terra rule within their twelve-test slices, with no adjusted finding across the full 144-test sensitivity grid. Two input-intervention batches totaling 5,184 responses supply paired intervention-return tests. The three-model batch has no bootstrap-adjusted finding at the primary block length; a Luna row-order effect appears under heteroskedasticity- and autocorrelation-consistent (HAC) adjustment within its three-test family but not in a pooled 24-test adjustment. Compact rules approximate aggregate rankings; return conclusions depend on comparison families, uncertainty methods and tie priorities.

Abstract: Post-training has been shown to significantly improve language models' performance on tasks with verifiable outcomes, including mathematical reasoning, software engineering, and computer use. However, whether the same approach can improve forecasting in financial markets is much less clear. Compared with tasks with verifiable outcomes, not only are realized returns noisy, but even what constitutes a relevant information set for making effective predictions is not obvious a priori: the model must decide which observations to gather and then commit to a numerical judgment before the outcome is known. We study this question in a chronological stock-price sandbox, where a language model gathers price, volume, relative-performance, and market-context evidence and predicts a future return. We post-train Qwen3-4B with supervised fine-tuning (SFT) on tool-use demonstrations, then proximal policy optimization (PPO) with a terminal reward given by the forecast score against the realized return. The resulting AURA-4B more than doubles the starting direction--magnitude score, from 20.94 to 43.31, and is comparable to frontier language models on this benchmark. Conditional magnitude agreement rises from 33.3 to 66.2, while directional accuracy changes from 62.9 to 65.4. SFT expands tool use, and PPO further increases the share of ranking and market-context queries. These results show that post-training can substantially improve financial forecasting performance, together with changes in how the model investigates the market, on this outcome-selected benchmark.

Abstract: 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. This is challenging due to several factors: financial markets exhibit non-stationary dynamics with shifting regimes, multimodal inputs such as price and news data often contain stochastic noise, and existing diffusion-based approaches, while effective for modeling stochastic dynamics, rely on assumptions such as isotropic Gaussian noise that fail to capture the state-dependent nature of financial uncertainty. To address these challenges, we introduce RADAR, a retrieval-augmented diffusion framework that learns market representations by conditioning on similar historical regimes. RADAR leverages retrieval to construct context-dependent noise distributions, applies conditional diffusion to denoise multimodal representations, and initializes the diffusion process using empirical statistics to reflect state-dependent uncertainty. Experiments show that RADAR achieves state-of-the-art performance on key risk-adjusted metrics while producing economically meaningful signals on asset returns and correlations.

Abstract: We present a three-tier progression of Hidden Markov Models: maximum-likelihood (Baum-Welch), variational Bayesian (VBEM), and a neural variational extension (Neural VBEM), that model each customer's transaction history as a trajectory through a small number of latent behavioural regimes, one of which is empirically identified as fraud-associated. The Neural VBEM HMM replaces the fixed Gaussian-multinomial emission family with a learned encoder, compressing a 741-dimensional transaction representation into a 64-dimensional latent space in which the VBEM HMM's posterior operates; a UMAP projection of this space reveals that the discovered regimes are not discrete clusters but ordered segments of a single continuous behavioural manifold, with confirmed fraud concentrated at its extreme. We show that the model's natural output, that is, the posterior probability of regime membership, is routinely mistaken for a fraud probability, and quantify the resulting miscalibration (the regime-membership interpretation error, MRIE); a corrected posterior-predictive score, closes most of this gap. We further distinguish batch (smoothed) inference, which uses look-ahead unavailable at deployment time, from filtered (forward-only) inference, and report both. On IEEE-CIS transaction data, the neural tier achieves a 14.4$\times$ fraud enrichment in its identified regime; while its AUPRC trails a discriminative XGBoost baseline, we show this gap is structural and not incidental, and argue the model is best positioned as a calibrated triage and interpretability layer rather than a drop-in ranking replacement.

Abstract: 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. Long-horizon alpha discovery provides a state-dependent setting: once a new factor enters the portfolio, the predictive information already covered changes, so the value of the same candidate or experience may change over time. We introduce EverMine, an empirical framework for studying self-evolving research capabilities in long-horizon alpha discovery. EverMine decomposes the research state into history (Hist), the current factor portfolio (Frontier), and reusable capabilities (Cap). Under matched resource limits, we compare complete runs with fixed or evolving Cap, and replace Cap while holding Hist and Frontier fixed to estimate the conditional value of accumulated capabilities. We also combine full trajectories with historical-state replay to examine how experience-based decisions affect candidate selection and portfolio outcomes. Across 18 long-horizon trajectories, end-to-end comparisons show no consistent gain from Cap evolution. Across 48 continuation branches from shared Hist and Frontier states, accumulated Cap also does not consistently outperform the initial Cap. Parameter tuning of existing factor structures can still improve the portfolio. In an exploratory replay of two screening batches from one Evolving trajectory, some screened-out candidates have positive marginal value at the original state, yet submitting all screened-out candidates sequentially slightly lowers final portfolio IC in both batches. These results show that candidate value depends on the evolving portfolio and submission order, and motivate evaluating self-evolving research capabilities through end-to-end outcomes, conditional capability value, and the consequences of experience-based decisions.

Abstract: 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. Our answer is governed self-evolution: the agent may propose, and a frozen statistical referee that the agent cannot touch must judge. The referee scores each candidate only on market outcomes revealed after submission, by betting, so its false-discovery guarantee holds at every stopping time for any proposal policy. We cross three proposers (a script, a bandit and a language model) with this referee and with three deliberately leaky ones, in a synthetic world with planted truth, a probe-authoring environment and a ten-year walk-forward on the CSI 500. Who judges sets the number of false admissions: the frozen referee admits 5-11 times fewer sub-threshold factors than the leaky referees under a scripted proposer, and no proposer closes that gap. Who proposes sets the yield: the language model beats the script, matches the bandit, and adds the one capability a bandit lacks, writing its own diagnostic probes. The certificate's price is time: an admitted true factor waits about 500 trading days, and the certified portfolio's Sharpe ratio therefore trails an ungated one. Judging belongs to the procedure; proposing and instrument-making belong to the agent.

Abstract: 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. Computer science research has developed strong methods for time-series forecasting, text classification, multimodal stock prediction, graph-based market modeling, and machine-learning operations, yet these streams do not provide a domain-specific protocol that jointly tests financial language-model outputs under event-time observability, probability calibration, execution timing, transaction costs, liquidity constraints, capacity limits, operational diagnostics, and statistical inference. We introduce MFAST, a Market-Friction-Aware Sentiment-to-Trading framework that converts timestamped financial text into auditable, reproducible, and market-feasible trading decisions. The application is news-based trading, where firm-specific text must be linked to securities before portfolio decisions can be evaluated. The framework links Refinitiv News Analytics to Center for Research in Security Prices (CRSP) equity data, restricts the primary out-of-sample evaluation to post-release news outside disclosed foundation-model data-freshness periods, and adds a public replication arm using open financial text and public price data. Results show that decoder-only language models outperform encoder baselines and dictionary sentiment in classification, calibration, return prediction, and net portfolio performance, while operational diagnostics reveal trade-offs among accuracy, latency, memory, throughput, and inference cost. The paper shows that credible evaluation of financial language models requires an end-to-end engineering approach combining language understanding, temporal discipline, market-friction-aware deployment, and reproducible validation.

Abstract: Hybrid Deep Learning for equity index forecasting is limited by three problems: propagation of OHLCV noise into derived technical indicators (TIs), channel-indiscriminate multi-scale decomposition that conflates heterogeneous frequency signatures, and static multi-branch fusion that cannot adapt to market regime shifts. WaVeFuse addresses these limitations through a unified dual-branch architecture. Symlet-4 wavelet denoising (level 2, MAD soft threshold) suppresses microstructure noise in OHLCV. Seven low-lag TIs computed from denoised prices are encoded by a causal channel-wise continuous wavelet transform (Morlet, 32 scales) into a per-timestep scale-space matrix. A CNN-BiLSTM branch captures temporal dynamics, while a dual-layer Transformer (heads=4, dk in {16, 32}) models inter-scale spectral dependencies, and their representations are integrated by a 2-token softmax gate Vertical Attention Fusion (VAF) that dynamically reweights branches as market regimes shift. Evaluated under walk-forward validation (WFV) on KOSPI, DAX, NYSE Composite, and Russell 2000 (2010-2023), WaVeFuse achieves R2 = 0.81-0.96 and directional accuracy 70.5-78.3%. It outperforms seven state-of-the-art models by 8.9-20.2% MAE across twelve dataset-period configurations. Diebold-Mariano statistics (4.62-10.38, p<0.001) confirm superiority over a well-tuned XGBoost benchmark across four indices. Ablation verifies component-wise contributions. Under realistic backtesting with 10 basis point transaction costs, WaVeFuse's directional strategy achieves a mean Sharpe ratio of 3.69 across four markets and limits maximum drawdown to 7.5% during the COVID-19 crash. With 152k parameters (0.68MB) and sub-1.3ms GPU inference, WaVeFuse delivers a computationally efficient, regime-robust framework suitable for research and decision-support deployment.

Abstract: Reinforcement learning trading systems published in the academic literature overwhelmingly rely on price-aggregate state representations (OHLCV bars) or limit-order-book depth features, leaving microstructure pattern theories from the practitioner literature, namely Auction Market Theory and Market Profile, without a peer-reviewed computational instantiation. We present ViperQ, a reinforcement learning system whose state representation is built explicitly from Auction Market Theory primitives: Volume Point of Control, Value Area position, Low Volume Node flags, Cumulative Volume Delta divergence, and tape-velocity signatures, assembled into a 20-dimensional Z-normalised vector. Two Proximal Policy Optimisation agents are trained with a prospect theory-grounded asymmetric reward function that penalises losing holds at a magnitude consistent with Kahneman and Tversky's loss-aversion coefficient. Evaluated on a held-out twelve-month partition of institutional tick data the agents have never seen, ViperQ achieves +163.6% ROI on TSLA (-27.5% max drawdown, 27,019 trades) and +116.5% ROI on NVDA (-47.8% max drawdown, 12,892 trades) under zero leverage. The results establish Auction Market Theory features as a tractable structured input modality for sequential decision-making on financial time series and motivate further work on microstructure-aware policy learning.

Abstract: Financial NLP has a standard workflow: validate a sentiment tool against human labels, then trust it to extract market signal. This assumes the two evaluations measure the same thing. We test that assumption in a setting where both can be measured at once: a corpus of securities class actions (2002-2025) linking 70,500 X messages to abnormal stock returns, with a single-annotator human labelled gold sample. Running five instruments (VADER, Loughran-McDonald, FinBERT, Twitter-RoBERTa, and an LLM annotator) through one identical pipeline, we find that the relationship between construct and predictive validity depends on the sampling convention and score representation. Under conventional method-specific sampling, human agreement aligns more closely with graded same-day associations than with one-day leads. On a fixed-n panel, however, agreement has similar graded rank correlations at both horizons, while the coarse ordering remains weak. Benchmark agreement therefore establishes semantic validity but does not by itself determine predictive rankings. In a conversation that is 17.6% spam, message volume predicts neither market damage nor settlement size.

Abstract: 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.

Abstract: MASTER's inter-stock multi-head attention -- the module responsible for modeling cross-sectional stock relationships -- accounts for 42.5% of model parameters and 25% of predictive value. We systematically decompose this module and uncover a surprising structure: the learned attention is near-uniform (perplexity 278/300), yet forcing exact uniformity eliminates all cross-sectional discrimination. Spectral analysis resolves this paradox: the deviation from uniformity is low-rank (effective rank ~65, top-10 modes capture 96.5% of energy), explaining why sparse approximations consistently fail while Nystrom low-rank attention (m=32 landmarks) matches full O(N^2) attention at O(mN) cost -- certified equivalent via TOST at both N=300 (5 seeds, Rank IC p=0.003) and N=800 (10 seeds, Rank IC p=0.034). Additional findings include: (i) attention anti-correlates with return similarity (Spearman rho = -0.614; on the industry-labeled subset, -0.645 unconditionally and -0.627 after controlling for industry, beta, and volatility), suggesting complementarity-seeking rather than correlation mining; (ii) all graph-based alternatives degrade performance, with hard masking worse than complete module removal; and (iii) at N ~ 3,500 with adapted architectures, no cross-stock module (GCN, Nystrom, or MASTER-style pipeline) significantly outperforms a per-stock LSTM baseline (n=4 seeds), indicating that the benefits observed at smaller scales do not trivially transfer. These results establish that the inter-stock attention's value resides in a compressible, dynamic, near-global redistribution that rewards low-rank approximation but resists sparsification.

Abstract: Cross-sectional return prediction from raw intraday bars is sensitive to each instrument price level, an additive nuisance under a return-ranking hypothesis. We test whether removing this offset, rather than rescaling amplitudes or changing the encoder, explains gains on a point-in-time CSI 300 five-minute panel. Eight parameter-matched encoders are evaluated with and without RevIN normalization; a parameter-free ladder then separates identity, scale-only, centering, last-value referencing, differencing, and standardization across all fields and restricted channels. Centering drives the reliable effect, while scale-only normalization does not help. All eight paired effects are positive and survive Holm correction on raw rank IC, after style residualization, and after additionally residualizing on short-term reversal. Among six stronger encoders, normalized IC is 0.0830-0.0939 and gains are 0.0376-0.0567. Price-only standardization retains 93-101% of the all-field gain. These results place the main effect in transformed price-channel offset removal rather than amplitude scaling or encoder choice.