Large language models systematically over-extrapolate recent trends when asked to forecast, and prompting cannot fix it. We propose a supervised fine-tuning approach using Low-Rank Adaptation (LoRA), training models on instruction datasets built from rational benchmark forecasts to intervene at the parameter level in how LLMs translate observed information into predictions. In controlled forecasting experiments and stock return prediction tasks, a single round of low-cost fine-tuning aligns AI forecasts with rational benchmarks out-of-sample, correcting extrapolative bias while maintaining generalizability.

We develop a statistical test to detect lookahead bias in economic forecasts generated by large language models (LLMs). Leveraging state-of-the-art pretraining data detection techniques, we estimate the likelihood that a given prompt appeared in an LLM's training corpus—a statistic we term Lookahead Propensity (LAP). We formally show that a positive correlation between LAP and forecast accuracy indicates both the presence and magnitude of lookahead bias. We apply the test to two forecasting settings: news headlines predicting stock returns and earnings call transcripts predicting capital expenditures. Our approach provides a cost-efficient diagnostic tool for assessing the validity and reliability of LLM-generated forecasts.

Large language models pretrained on internet-scale data risk lookahead bias in forecasting tasks, as they may have already seen the true outcome during training. To address this, we present DatedGPT, a family of twelve 1.3B-parameter language models trained from scratch on approximately 100 billion tokens each with strict annual data cutoffs spanning 2013 to 2024, together with DatedInstruct, an instruction dataset grounded in each year's documents to prevent leakage during post-training. The models are competitive with open models of similar scale, and perplexity-based probing confirms that each model's knowledge is bounded by its cutoff year. On stock return prediction over 61,000 firm-day news headlines, DatedGPT-instruct achieves an annualised Sharpe ratio of 3.20 under the lookahead-bias-free setup. Lookahead-biased models, whose training data covers the outcome period, add a lookahead premium of 26.4 b.p. per standard deviation, significant at the 1% level. The series thus enables direct analysis of lookahead bias in financial forecasting.

We employ Natural Language Processing (NLP) to scrutinize regulatory filings, identifying institutional investors' climate change preferences and risk perceptions. These preferences and risk perceptions grow over time and are stronger after a fund has signed the Principles for Responsible Investment (PRI) or is located in regions with stronger global warming beliefs. Investors preferring green assets tend to decrease their portfolio weights in environmentally unfriendly stocks, reflecting a desire to align investments with their values. However, the relationship between climate risk perceptions and portfolio weights of brown stocks varies due to heterogeneous investment strategies. Investors with higher climate risk perceptions are more likely to support environmental shareholder proposals, whereas investors with green preferences are not. These findings provide new insights into sustainable investing behavior under differing investor motivations.

A multi-armed bandit algorithm is developed for factorial experiments. Using tools from advanced probability theory, I first prove that UCT algorithm with Laplace bound has a lower computational complexity than the naïve UCT algorithm. I begin by analyzing UCB1 for non-stationary bandit problems, and then prove UCT algorithm with Laplace Bounds achieves a better lower bound. Also, I demonstrate that the probability of suboptimal choices will converge to zero with a convergence of failure probability. In settings of deep learning, experimental results are also consistent with the theoretical regret bound.

Using Factorial experiments, I explore multi-armed bandit problems in which a player selects actions (here a sequence) episodically and observes the outcomes. I consider the Upper Confidence Bound applied to Trees (UCT), a popular algorithm for tree search, in order to identify the sequence of choices that maximizes some objective function. Using synthetic experiments, I demonstrate that applying tighter concentration bounds to Linear Bandits can significantly improve the performance of UCT for tree search. Next step is to investigate various factorial experimental design configurations. I also compare the performance of algorithms under three different formulations of the factorial experiment: 1) standard bandits; 2) linear bandits; and 3) bandits for tree search. I observe that capturing the underlying tree structure is essential for robustness, whether the outcome function is linear or not. Furthermore, I observe that the algorithms employed under the bandits formulation of tree search for factorial experimental designs appear more robust to the noise variance than other approaches.