TL;DR
It is concluded that the genuinely difficult part of deployment is tail-event detection rather than directional forecasting, and that a credible evaluation must rely on event-sensitive metrics and time-aware validation; without them, performance under heavy imbalance is easily overstated.
We present a language-model forecasting system for merger arbitrage, a specialized high-stakes financial setting in which the task is to predict the outcome of announced M\&A deals. Unlike prior work on judgmental forecasting with LLMs, which has focused on broad mixed-topic benchmarks and short context such as news snippets, we study a setting that requires long-context reasoning over hundreds of pages of technical documents. Our system combines expert-guided context engineering with finetuning on hindsight-guided reasoning traces derived from historical deals. Given an announced deal, it outputs a probability distribution over three mutually exclusive outcomes: closing at announced terms, a higher bid, or deal termination. On an out-of-sample set of more than 400 large deals spanning 42 countries, our finetuned system achieves the best performance of any method we evaluate, reducing class-balanced Brier score to 0.151. This is 24\% below calibrated market-implied probabilities, 19\% below XGBoost, and 25-42\% below frontier language models. These results, together with ablation studies, show that LLM-based forecasting can succeed in specialized, long-context financial workflows, with hindsight-based supervision and expert-designed context playing a critical role.
Large pretrained time-series models such as TimesFM are attractive for financial forecasting, but raw directional accuracy is a misleading scoreboard in equity markets. An early LoRA adapter in this project appeared to reach roughly 80% directional accuracy; we show this is not evidence of skill. Over a long horizon in a rising market, a trivial"always-up"rule attains comparably high accuracy without using the input at all. To separate genuine skill from this base-rate artifact, we build a reproducible, frozen-data benchmark with expanding walk-forward folds, a stratified held-out-ticker split, honest baselines (zero-shot TimesFM, always-up, random-walk, persistence, AR(1)), and paired significance tests (McNemar, Diebold-Mariano) under Benjamini-Hochberg FDR control. We apply the identical method to two universes -- a tech-heavy NASDAQ-100 and a broad S&P 500 -- reporting excess accuracy over the always-up base rate. Three findings replicate. First, when the historical ~80% condition is recreated, the high number is a base rate of ~0.70 that the fine-tuned model scores below. Second, pooled LoRA shows no directional skill over the base rate at any horizon on either universe (negative at the six-month horizon). Third, per-sector specialization is significantly worse than a single pooled adapter (Diebold-Mariano p<0.001 on held-out stocks at h=128). Fine-tuning's only measurable benefit is a statistically significant reduction in point-forecast error relative to zero-shot TimesFM, which nonetheless does not beat naive baselines and confers no tradeable directional edge. The contribution is methodological: a defensible, fully seeded protocol that prevents the base-rate trap, together with the replicated negative result it produces.
This work examines the pipeline on Russell 2000 equities under three stock-selection regimes and suggests that stock-selection regime and allocator choice matter at least as much as the sentiment model, and that separating firm-specific and macro-exposure triggers is more informative than requiring both to fire simultaneously.
Predicting financial asset returns remains one of the most difficult challenges in empirical finance, driven by the low signal-to-noise ratio and the semi-strong form of market efficiency. While deep learning models, especially LSTM networks, have shown promise in capturing temporal dependencies, standard architectures often struggle to account for the cross-sectional heterogeneity of asset returns. This paper proposes a novel architectural extension to the basic LSTM model designed to improve both predictive accuracy and model interpretability. The framework integrates macro-financial covariates to capture broader economic signals and learnable sector embeddings to encompass heterogeneity by sector. The trading strategy involves constructing a long-short portfolio based on daily directional forecasts for each S&P 500 constituent, targeting stocks expected to under- or outperform the cross-sectional median return of the S&P 500. Model Performance is evaluated against three competitive benchmarks: a basic LSTM, a Random Forest model and a traditional market buy-and-hold strategy. The empirical results demonstrate that the LSTM with sector embeddings outperforms all benchmarks across key risk and return metrics. By utilizing sector embeddings, the model explicitly incorporates cross-sectional heterogeneity, allowing it to adapt to varying industry dynamics within the market. To address the black-box nature of deep learning, I use latent space visualizations to analyse how the model differentiates between sectors, providing insights into the internal representation of the sectors in the LSTM. The impact of the sector information can be quantified using a novel contribution metric by inspecting the weights of the LSTM. The predictive signal is driven by a short-term reversal factor and an industry momentum factor.
Leveraged foreign-exchange trading is exposed to rapid adverse price movements that can contribute to margin-call events, yet most prior studies emphasize price or direction forecasting rather than early risk classification. This study uses the Sample, Explore, Modify, Model, and Assess (SEMMA) framework to organize a time-aware experiment on hourly EUR/USD and GBP/USD data from 2010 to 2026. After cleaning, the datasets contained 99,987 and 99,985 observations, respectively. A direction-agnostic maximum adverse excursion proxy labeled whether either a hypothetical long or short position would experience at least 50 pips of adverse movement within five hours. Sixteen technical features were converted into 60-step sequences and evaluated using identical five-fold chronological walk-forward splits. XGBoost, BiLSTM, Attention-BiLSTM, TransformerEncoder, and PatchTST-Lite were compared using training-only scaling and validation-only threshold selection. XGBoost achieved the strongest mean F1-score and ROC-AUC on EUR/USD (0.3710 and 0.7927) and GBP/USD (0.4855 and 0.7847). Attention-BiLSTM remained competitive, with F1-scores of 0.3583 and 0.4825 and the highest mean recall on GBP/USD (0.6680). In a no-retraining EUR/USD-to-GBP/USD transfer test, it obtained an F1-score of 0.5152 and ROC-AUC of 0.7923. Five-fold Wilcoxon tests lacked sufficient resolution to establish superiority. The results support Attention-BiLSTM as a temporally attributable early-warning component, while XGBoost offers the best efficiency-performance trade-off.
Accurate time series forecasts underpin asset pricing, risk management, monetary and macroprudential policy, and other applications. The set of available forecasting methods is expanding rapidly, driven by new machine learning models. This raises a practical question: Do these methods deliver real forecasting power gains on financial data? Since no single method is best across all data generating processes, the question can be answered only by direct evaluation on domain-specific data, and a fair comparison requires holding the data fixed across methods. Yet, new methods are rarely evaluated on the financial data commonly used in academic finance; when they are, each is assessed on its own dataset and cleaning conventions. Apparent method rankings entangle method skill with cleaning choices that themselves require domain expertise. We address this by assembling, in one place, canonical cleaning procedures for many financial datasets to hold the data fixed across forecasting experiments. We introduce a standardized open-source dataset covering equities, corporate bonds, U.S. Treasuries, foreign exchange, commodities, credit default swaps, options, five basis spread datasets, and bank and intermediary indicators, each cleaned per the canonical paper for that asset class. Holding the data fixed and evaluating roughly a dozen univariate methods without exogenous regressors, we find that asset returns remain near unforecastable across every method family and that hybrid and machine learning methods exhibit additional forecasting power on basis spreads and bank indicators.