Text encoders turn earnings calls into vectors for search, for language models that answer from retrieved passages, and for forecasting models. A live retriever can use an encoder trained once on all available text, as finance retrievers are, but a backtest needs an encoder that read no text written after the date of each vector. To address this look-ahead bias, from 412,248 calls we train vintages under four pretraining recipes, with yearly cutoffs from 2010 to 2024, and each vintage reads only text dated on or before its cutoff. We then fine-tune each vintage on the question-answer pairs of the calls themselves. The vintages we train on calls alone, with 33.8M parameters, retrieve answers on held-out calls better than the general retrievers we tested in every test quarter. Their vectors also raise by about half the rank correlation between next-day returns and forecasts built from past returns and firm size. When each vintage continues the previous one, a later vintage retrieves better on earlier quarters than the vintage available at the time, and a point-in-time language model that answers from the passages the later vintage retrieves inherits the gain. The 2024 vintage of our families also forecasts earnings surprises better, although the interval of this gain includes zero on calls that no vintage read. It forecasts returns no better, so point-in-time vintages matter most for retrieval and little for forecasts of returns.
This technical note introduces tabular foundation models (TFMs) to accounting research and evaluates their performance in one common benchmark task: predicting firms' next-year earnings changes. TFMs produce numeric predictions and learn in context, without post-release training or tuning, making them natural and easy to use for forecasting. The TFMs we study are pretrained on synthetic tabular data, so they avoid look-ahead bias from memorizing firms' outcomes. We compare Google Research's TabFM with a fine-tuned open-weight large language model developed in recent accounting research and with several tree-based models. TabFM is more accurate than all comparison models. It is also faster and less computationally intensive than the large language model. Several tree-based specifications produce predictions more quickly than TabFM, so choosing between them involves a trade-off between accuracy and speed. We conclude that TFMs have promise as a new tool for forecasting in accounting and finance. However, our evaluation is based on a single benchmark task, and no class of models is likely to dominate in all settings.
A hedging objective reflects the risk preference of the hedger, so the conventional approach is to estimate the hedge ratio directly under the chosen objective. This paper asks whether that estimate should be used alone. We construct five objective-specific rolling hedge-ratio estimators---minimum variance, least absolute deviation (LAD), downside semivariance, lower-tail quantile, and Expected Shortfall (ES)---from the same historical stock--market returns, and combine the resulting estimates under a fixed final evaluation objective. Separate multi-output ExtraTrees models predict the within-stock-date ranks of the five candidates under MSE, LAD, downside semivariance, and 10\% quantile loss. The design uses pre-specified 60-, 90-, 120-, and 180-day windows with objective-matched rolling estimators as benchmarks. Adaptive weighting improves the corresponding loss in all sixteen objective--window comparisons. Equal weighting, which uses the same objective-specific hedge ratios without state variables or learned reweighting, already improves downside and quantile hedging. The results therefore suggest that estimates generated under other objectives can be useful even when the final hedging preference and evaluation objective are fixed, while adaptive weighting can provide a further improvement.
Quantifying the uncertainty of forecasting models is essential to assess and mitigate the risks associated with data-driven decisions, especially in volatile domains such as electricity markets. In this paper, we propose a computationally efficient isotonic variant of Quantile Regression Averaging (QRA), called Isotonic Quantile Regression Averaging (iQRA), for generating probabilistic forecasts from ensembles of point forecasts. Building on QRA, we introduce stochastic order constraints as a hyperparameter-free regularization mechanism that reduces computational complexity while maintaining competitive forecast accuracy. We further demonstrate that the same isotonic regularization principle can be applied to the smoothed estimator of QRA. In an extensive forecasting study of two European day-ahead electricity markets, we show that iQRA and its smoothed variant consistently rank among the top state-of-the-art postprocessing methods in terms of reliability and sharpness. It produces well-calibrated prediction intervals, with smoothed iQRA providing superior reliability even when compared against coverage-oriented conformal prediction. We further introduce a battery arbitrage evaluation, showing that the leading probabilistic methods achieve the highest profit per trade and that probabilistic risk control improves the economic value of trading decisions. Isotonic regularization offers a scalable hyperparameter-free variable selection for linear quantile regression, making iQRA attractive for practical decision-support systems in energy markets.
High-frequency volatility forecasting typically separates measurement from prediction: intraday observations are first compressed into a daily realized measure using a fixed aggregation rule, and the resulting statistic is then passed to a forecasting model. This conventional pipeline therefore treats measure construction as forecast-agnostic preprocessing. We ask whether the aggregation stage itself can instead be learned from the downstream forecasting objective while preserving the statistical structure of established realized measures. We propose Neural Realized Measures (NRM), a primitive-preserving, forecast-supervised aggregation framework that retains the original measure-specific intraday contributions but replaces uniform aggregation with positive, mean-normalized context-dependent weights. Uniform weights exactly recover the corresponding raw realized measure (RRM), so the classical construction is nested as a special case. Because NRM remains an explicit weighted sum of established primitive contributions, it also retains direct bar-level attribution. Across SF, HAR, and MIDAS forecasting specifications, NRM is evaluated against corresponding RRM specifications under controlled comparisons. In Heston simulations, NRM reduces RMSE for both contemporaneous and next-day latent integrated variance across all evaluated primitive--regime combinations. On six cryptocurrencies using five-minute data, NRM improves QLIKE in all 54 paired out-of-sample comparisons, with mean gains of 7.23\%, 8.15\%, and 9.22\% under SF, HAR, and MIDAS, respectively. Validation-selected NRM specifications also achieve the lowest mean QLIKE across the six assets among the evaluated approaches. Ablation and SHAP analyses reveal systematic time-of-day and nonlinear local-context effects in the learned allocation.
Under temporal nonstationarity, model performance rankings reverse across training horizons due to heterogeneous sensitivities to memory. While existing literature addresses adaptive lookbacks, effective model complexity, and empirical forecast evaluation in isolation, no unified framework explains the structural mechanics driving these rank reversals. We bridge this gap by deriving a model-agnostic decomposition of out-of-sample $R^2$ into target alignment, miscalibrated aggressiveness, and predictive instability. Within this framework, we mathematically characterize complexity phase boundaries, establish closed-form optimal-memory equations for ridge regression under local drift, and extend these guarantees to complex model architectures via surrogate bounds. We validate these theoretical dynamics by benchmarking 14 linear, tree-based, and neural architectures across large-scale financial panels. In doing so, our framework demonstrates that empirical model rank reversals are driven by predictable surges in miscalibrated aggressiveness, providing a rigorous mechanistic diagnosis of generalization under temporal shift.
We develop PanelGP, a conditional factor model for large panels of stock returns in which firm characteristics determine the factor loadings and economic conditions determine the expected factors. Both relations are estimated nonparametrically with Gaussian-process priors. The kernel choice nests instrumented principal component analysis (IPCA) as well as smooth nonlinear maps. Asset-specific loading innovations and common factor innovations separate idiosyncratic from systematic risk, so one estimated model forecasts both expected returns and their covariance matrix. Estimation uses sparse variational inference with a joint posterior over loadings and factors. Because each predicted return is bilinear in a Gaussian state, the predictive covariance matrix and the cumulants of any portfolio have closed forms. In simulations, the estimator recovers the nonlinear maps and the variance parameters. For S&P 500 constituents over 1995-2024, nonlinear loadings rank stocks better than affine loadings and improve the Gaussian log score. Relative to IPCA, the resulting portfolios have higher Sharpe ratios, also net of trading costs, and the minimum-variance portfolio has lower volatility. The predicted variances rank stocks and months by risk but are too compressed.
Large language models (LLMs) read company news at scale, and backtests suggest that their readings predict stock returns. We ask how much of that predictability survives a test that rules out look-ahead, memorization and selection. We propose OLIVIA, a protocol for such tests: point-in-time inputs with company identities masked, selection on validation data only, an analysis plan frozen by cryptographic hash, and a test period unblinded once through a logged lock. We apply it to AI decision models, non-generative language models that return probabilities for typed questions, reading news about S&P 500 stocks: an open-weights model we fine-tuned on 2024-2025 data, three commercial models and standard baselines. The fine-tuned model's signal replicated on two validation quarters (daily information coefficients 0.055 and 0.065) but fell to 0.020 (t = 1.7) on the blinded January-July 2026 test. The commercial models' readings line up with the overnight move to the open (information coefficient 0.074 to 0.077) but not with the move after it, so the news is priced by the time a trader can act on it at the open. No model can tell over- from under-reaction, and no pre-registered strategy survives costs and a deflated Sharpe ratio. Commercial models identify masked companies in 62% to 74% of cases against a 28% baseline: masking hides names, not companies. The market reads the news first.
Quantum reservoir computing (QRC) processes temporal information through quantum dynamics, but it remains unclear whether small QRC models can forecast real data competitively with classical models and whether their predictions survive the transfer from ideal simulation to noisy hardware. We address both questions for a six-qubit gate-based reservoir that forecasts next-day motor-insurance claim counts. Five reservoir candidates were selected and frozen using hardware-native interaction graphs, Trotterized Ising dynamics, memory and measurement-robustness criteria, circuit cost, and calibrated-noise screening. On 232 held-out test days, the ideal candidates reached root-mean-square errors (RMSEs) of 5.66-6.12 claims/day, compared with 5.36 for ridge regression. For four of the five, the difference from ridge was not significant under a 95% block bootstrap, and their accuracy was comparable with classical recurrent and reservoir baselines. On the 156-qubit IBM Kingston processor, we ran 63 evaluations in 21 paired rounds, comparing hardware-aware, random-native, and Qiskit optimization level 1 layouts. Hardware-aware placement gave lower RMSE than Qiskit in every round, with a maximum transfer distortion of 7.59 claims/day (11.84 for Qiskit), and avoided the severe transfer failures seen for some random layouts. Most of the hardware distortion appeared as a systematic forecast shift, which a short hardware probe on training inputs detected before the test runs. Longer replay windows raised circuit cost without a consistent gain in forecasting accuracy or hardware transfer. Small QRC models can thus reach competitive accuracy on this task, and hardware-aware layout selection makes their transfer to IBM Kingston more consistent.
Financial workflows need fast, calibrated, auditable decisions, not just generated text. We train System-1 decision models, small language models fine-tuned to answer typed questions (choice/score/noul) with a probability in a single forward pass, across two model families (K2-Horizon and Qwen3.5/3.8) and nine rungs from 0.8B to 32B parameters, using one frozen LoRA recipe on a public decision benchmark and two new finance corpora: a programmatic-gold policy corpus (rule-reading with exact labels) and a reality-labeled outcome corpus (19,640 realized forward-return questions in walk-forward splits, plus a sealed 2026 test of 1,360 questions). Narrow fine-tuning costs general decision competence 9-45 percentage points, a data-mixture effect that shrinks with model scale and is associated with narrow-gradient volume. On reality-labeled questions, every trained rung from 0.8B to 32B lands at the horizon-conditioned prior baseline (accuracy 0.37-0.38, no detectable conditional edge under our controls); the models respond with flat, well-calibrated probabilities that track per-horizon base rates: naturally honest uncertainty, with a mid-scale calibration sweet spot (ECE 0.026 at 2B, degrading to 0.043-0.044 at 9-32B under prior drift). A mixed-corpus remediation carries all competencies in one adapter at modest cost: it matches or slightly exceeds the generalist specialist on the public benchmark while holding the policy ceiling and near-specialist outcome calibration, yielding a deployable artifact with per-domain temperature regimes and a band-dispatch escalation analysis: sharp, dispatchable bands where rules determine answers; flat, review-routing bands where reality does not.
Financial research agents can generate and monitor trading rules faster than a human can document the resulting selection process. Candidates appear at data-dependent times, weak rules are abandoned, and promising ones are watched until they cross a reporting threshold. We introduce PROMOTE, a protocol that assigns every candidate predictable testing wealth at birth and accumulates a cost-aware e-process only from subsequent returns. If lifetime allocations sum to at most one, a weighted Ville argument bounds the probability of ever promoting a rule whose declared bounded net score has always had nonpositive conditional mean. The result permits dependent candidates, adaptive births, and indefinite monitoring. We audit the implementation in two settings. A retrospective pseudo-out-of-sample replay contains 24 panel--rule hypotheses, 3,000 shared-sign null paths, injected alternatives, and public ETF and industry data. Repeated unadjusted tests produce a family crossing in 94.5% of null paths, versus 0.0% for PROMOTE; at a 15-bp aligned ETF-component perturbation, PROMOTE has 70.1% target power and median crossing session 2,327. A separate exact mean-zero experiment lets an agent select candidate identities and birth times from past scores; the funded gate has 1.77% false-family crossings across 3,000 programs. The market replay produces zero promotions after costs. The certificate controls historical false promotion under the declared null; it is not a certificate that an earlier edge remains profitable at deployment. AI-use disclosure. Generative AI tools, including OpenAI Codex, assisted manuscript preparation. For this preprint version, Codex assisted with LaTeX editing, author and venue metadata, compilation, and layout checks.
A model pretrained to predict a customer's next event can estimate the probability of future events without training on their labels and provide a reusable representation for supervised downstream banking tasks. We pretrain EventGPT, a hierarchical causal transformer, by next-token maximum likelihood estimation on the public Multimodal Banking Dataset (MBD-mini: 100,224 clients, 38.8 million transactions and 66.3 million geolocation pings). Keeping one state per event, EventGPT trains in about two hours on one L4 GPU at windows up to 8,192 tokens. From its next-event distribution, with no labels, EventGPT scores five targets of the form "the client makes a transaction of a given kind for the first time next month" at ROC-AUC 0.79 to 0.92. Gradient boosting on 114 engineered features needs 10,000 to 300,000 labeled client-months to match it on four targets and stays below it on the fifth. For product prediction, frozen EventGPT embeddings score below the engineered features (mean ROC-AUC 0.787 against 0.808 over five test folds), and a rank-averaged ensemble of EventGPT, profile fields and token counts with the engineered features adds 0.014 to 0.017, of which 0.007 to 0.016 is EventGPT's own. The code and pretrained weights are released.
The limit order book (LOB) records outstanding buy and sell orders across multiple price levels. Its state evolves over time across market side, price level, and order attribute, naturally forming a multiway structure. However, existing LOB models typically flatten these dimensions into a matrix model, thus obscuring the existing structural relationships. Starting from first principles and the market-microstructure intuition of the LOB, we design TensorLOB to operate entirely in the tensor domain. TensorLOB uses mode-wise tensor contractions for feature extraction and directly contracts the learned representation with low-rank class-weight tensors for prediction, retaining the distinct roles of time, side, level, and attribute modes. Experiments on the FI-2010 benchmark show that TensorLOB achieves the highest accuracy, recall, and F1 score across all four evaluated horizons. This proof of concept highlights the potential of high-dimensional modelling of multi-way latent information intrinsic to financial markets, which has so far been largely overlooked.
We identify an empirical inference-compute frontier for limit order book prediction on FI-2010. A power law fitted to low- and mid-compute models extrapolates to an excluded MLPLOB-style mixer family whose work is $56$--$2{,}515$ times the largest budget in the corresponding fit. Across nine folds and five horizons, pooled $R^2$ is $0.934$ on the realized target frontier, with 59\% lower RMSE than constant endpoint extrapolation, and $0.731$ on all excluded operating points. Measured latency gives a weaker relationship and reorders model families. We therefore design FastBiNLOB to pack computation into dense temporal and feature mixing. H120 exceeds the published state-of-the-art macro-F1 on two of four horizons in every one of five seeds. Under our CPU/MPS protocol, H96 reduces median latency by 23.6\% relative to the W128 MLPLOB reference, and H120 by 60.2\% relative to a TLOB-inspired dual-axis attention reference.
Generating realistic financial time series is challenging as training data is often limited to a single historical path. With such scarce data, overfitting is hard to avoid, especially under adversarial training where a trained discriminator can memorize the training samples. To mitigate this, recent approaches train generators to minimize the discrepancy between untrained feature representations of real and generated time series. In these works, the feature maps are based on path signatures, which can fail to capture relevant time series properties at tractable truncation depths. In this work, we instead train generators by matching random convolutional features of real and generated time series. Existing random convolutional feature maps, such as Rocket and Hydra, have been shown to provide informative representations of real-world time series, but cannot supervise generative models because they are non-differentiable. We introduce SOCK (SOft Competing Kernels), a fully differentiable random convolutional feature map, suited to train generative time series models. We show that generators trained by matching random SOCK features outperform signature and diffusion baselines on most evaluation metrics across a wide range of small-sample financial datasets. We further demonstrate SOCK's expressiveness on two-sample hypothesis testing and time series classification tasks, where SOCK matches or outperforms existing unsupervised feature maps.
This study aims to determine whether the application of Deep Reinforcement Learning (DRL) as a specialized execution overlay can enhance pair trading in highly volatile cryptocurrency markets. Although classical implementations of the strategy have proven successful in traditional equities, they frequently exhibit rigidity and suffer from severe divergence risks when applied to high-variance environments. To construct a robust system, we developed a hierarchical "Filter-then-Rank" pair selection methodology and a custom "Fixed Risk, Adaptive Mean" execution model. The system employs a Proximal Policy Optimization (PPO) agent with a Long Short-Term Memory (LSTM) layer to govern execution decisions within strict deterministic risk management boundaries. Evaluated on 1-hour interval data from the Binance USD-M Futures market, the selected RL policy achieved an Out-Of-Sample performance that substantially outperformed the heuristic baseline, although the results varied considerably across five random seeds, with four of the five seeds exceeding the baseline in both CAGR and Sortino Ratio. A stationary circular block bootstrap test of the Agent's risk-adjusted outperformance yields p-values of approximately 0.06, which fall just short of the conventional 5% significance threshold. Ultimately, this study contributes to the quantitative finance literature by introducing a hybrid architecture that combines statistical arbitrage with DRL execution policies. Furthermore, it applies safe reinforcement learning via deterministic shielding, with results indicating that anchoring a neural policy to statistically robust boundaries mitigates severe divergence risks.
Reliable comparison of deep learning architectures for financial forecasting requires controlled evaluation that separates architectural effects from hyperparameter tuning, random initialization, forecast horizon, and asset characteristics. We present a five-stage benchmark using equal-budget Bayesian hyperparameter optimization, category-level configuration freezing, three-seed final training, held-out metric aggregation, and rank-based statistical validation. Nine architectures from Transformer, MLP/linear, convolutional, and recurrent families are evaluated with direct multi-step prediction from five OHLCV features across 12 assets (four cryptocurrency, four foreign exchange, four equity-index instruments), using chronological 70/15/15 splits of 30,000 hourly windows per asset at 4- and 24-hour horizons. The benchmark includes 918 runs (270 optimization trials, 648 final training runs). Across 24 asset-horizon points, ModernTCN has mean rank 1.333 and the lowest RMSE in 18 points (75.0%), followed by PatchTST (mean rank 2.000). Global ranking differences are strongly supported by an Iman-Davenport test (F(8,184)=101.36, p<1e-15), with 33 of 36 Holm-adjusted pairwise comparisons significant. Rankings are consistent across horizons for all 12 assets (rho>=0.683; Stouffer p=3.47e-10), while errors increase at the longer horizon. Architecture explains 99.90% of raw error variance and 68.33% after within-slot normalization excluding LSTM; seed variance contributes at most 0.04% (0.02% without LSTM). Mean directional accuracy is 50.08%. These findings provide a statistically validated comparison of nine architectures, identify ModernTCN and PatchTST as leading performers under the evaluated conditions, and offer a reproducible benchmark for multi-asset financial time-series forecasting.
Factor models are central to asset pricing and portfolio construction, but neural factors are often difficult to interpret, and their effect on portfolio risk is unclear. A sparse autoencoder (SAE) may address both concerns by describing daily returns across many assets with only a few active features. We ask whether this sparse representation produces stable financial factors and better-diversified portfolios. We study cryptocurrency returns, use equities as a secondary test, and compare the SAE with common benchmarks in a time-ordered evaluation. We find consistently lower volatility in SAE portfolios and link this result to lower correlations among the selected assets. However, the SAE does not outperform a simple contrarian strategy on risk-adjusted return, and its factor identities are not stable across training runs. Our contribution is to show that sparse coding can support portfolio diversification, while clearly separating this risk effect from claims about alpha or stable factor identities.
We present, to our knowledge, the first adaptation of Pauli Correlation Encoding (PCE) to quantum topological data analysis, reformulating Betti number estimation as a depth-efficient variational optimization over a compressed qubit register. From a Takens embedding and Vietoris-Rips filtration of S&P 500 returns, we extract combinatorial Laplacians and recast null-space counting as a continuous-PCE Rayleigh-quotient minimization with variational deflation, encoding $n_k$ simplex indices into $O(n_k^{1/\kappa})$ qubits with shallow, ancilla-free circuits. Because the resulting loss is rational rather than bilinear in the correlators, the barren-plateau bound of Sciorilli et al. does not transfer; empirically the gradient variance decays only polynomially, with no exponential barren plateau, over $n=4$-$12$ qubits. The classical stage matches ripser on all 190 sliding windows (2003-2010). On the real market Laplacians ($\beta_1=1$-$22$), PCE-VQE recovers no null vector from a random start, but warm-starting from a classical null-space surrogate recovers $\beta_1$ exactly on all seven windows tested ($n_k \le 259$), placing the obstacle in the optimization landscape rather than the encoding. Chronologically split classification gives in-regime ROC AUC 0.818, but out-of-distribution evaluation on the 2020 COVID shock and 2022 rate cycle (AUC 0.009, 0.515) shows the calibration does not generalize across crisis regimes.