KQ• CV

01 / Research

Forecasts should survive contact with real time.

My work connects econometric theory, machine learning, and text-as-data with a consistent question: what information would truly have been available when a forecast was made?

01

New Evidence of the Usefulness of Real-Time Sentiment Indices in Econometric and Judgement-Based Forecasting Models

Kaiwen Qiu, A. Kabiri, John Landon-Lane, and Norman R. Swanson

Job Market Paper · Journal Submitted · 2026

EconBERTReal-time dataSurvey of Professional ForecastersMacroeconomic forecasting
Abstract +

In this paper we construct a set of new narrative sentiment indices using textual data from a large corpus of U.S. economic and policy documents. The indices are designed for use in real-time forecasting contexts and are constructed using transformer-based language models and LLM-assisted annotation. They include a finance-domain transformer; an economics-tone model trained on LLM-generated annotations of Federal Reserve, BLS, and Census text; and a topic-conditioned narrative measure.

To determine the usefulness of these three indices for forecasting, we carry out a series of experiments in which their marginal predictive content is assessed. We first assess the predictive content of the indices in econometric models used for real-time U.S. macroeconomic and financial forecasting. We then assess whether our sentiment indices contain information that can be utilized to improve the quasi-judgemental forecasts reported in the United States Survey of Professional Forecasters (SPF). This is done by testing the rationality of SPF forecasts when conditioning on both real-time macroeconomic variables and our sentiment indices.

In the real-time forecasting experiment, the autoregressive benchmark attains the lowest mean square forecast error in 11 of 44 target-variable and forecast-horizon cases, while one of the eleven alternative models attains the lowest mean square forecast error in the remaining 33 cases. Among these 33 cases, 31 involve specifications containing at least one sentiment index and two involve the macroeconomic-variable-only specification. In addition, we find that SPF forecasts incorporate the information contained in our three contemporaneous economic narrative indices, but do not incorporate the information contained in two widely used consumer-sentiment survey indices. Thus, real-time econometric forecasts can be made more accurate by incorporating information from our new narrative indices, while judgemental forecasts can be made more accurate by utilizing information available in consumer sentiment indices, but not by utilizing information available in our new narrative sentiment indices.

02

Consistent Factor Estimation and Forecasting in Factor-Augmented VAR Models

John Chao, Yu Liu, Kaiwen Qiu, and Norman R. Swanson

2026

Factor modelsVariable selectionFAVARForecasting
Abstract +

In this paper we establish that conditional mean functions associated with h-step-ahead forecasting equations implied by factor-augmented vector autoregressions (FAVARs) can be consistently estimated when factor pervasiveness does not hold. In particular, we begin by stating a common assumption of factor pervasiveness in which all available predictor variables, excepting a negligible subset, load significantly on the underlying factors. We then establish that even when this assumption is relaxed, consistent factor estimation can be achieved if one pre-screens the variables and successfully prunes out the irrelevant ones.

Furthermore, using factors estimated in this manner when constructing h-step-ahead forecasting equations implied by FAVAR models enables consistent estimation of the conditional mean function of those equations. Conditional mean functions constructed using our procedure are consistently estimable in a wide range of situations, including cases where violation of factor pervasiveness is such that consistent estimation is precluded in the absence of variable pre-screening.

03

Diffusion Indices, Policy Rules, and LLM Narratives: Forecasting the U.S. Yield Curve in Real Time

Kaiwen Qiu

2026

Yield curveLLM narrativesMonetary policyRDNSReal-time forecasting
Abstract +

Econometricians increasingly rely on machine-learning tools to summarize information in data-rich environments, yet the relative forecasting value of macro-based diffusion indices, text-based LLM sentiment indices, and monetary-policy-rule information remains unclear within a unified term-structure framework. This paper conducts a horse race across these information sets by embedding them in a dynamic Nelson-Siegel class model, including its rotated short-rate-based representation (RDNS).

We compare three families of real-time diffusion indices constructed from large macro panels: principal components (PCA), targeted shrinkage via LASSO and Elastic Net (EN), and the Chao-Swanson (CS) variable-selection approach. We further introduce three LLM-based text indices designed to capture policy-relevant narratives: an off-the-shelf FinBERT sentiment measure, an economics-domain-trained BERT index, and a topic-fusion fine-tuned index that integrates topic structure with supervised sentiment signals. Finally, we incorporate monetary-policy information through Taylor-rule components and Taylor-rule residuals, interpreted as policy shocks and motivated by the forward-looking policy-rule literature.

Using a real-time rolling-window design, we forecast yields across maturities and horizons. Our results show that CS-based diffusion indices extract more forecasting-relevant macro information than conventional LASSO/EN targeted-PCA constructions; LLM-based indices contain incremental predictive content beyond macro diffusion indices, particularly around policy-relevant episodes; and adding Taylor-rule components and/or policy shocks improves short- and intermediate-horizon yield forecasts, with the largest gains concentrated at the short end of the curve. Overall, combining policy-rule information with data-rich indices delivers the most reliable improvements in predictive accuracy for the U.S. Treasury yield curve.

04

Algorithmic Trading by Reinforcement Learning in a Collaborative Manner

Li Long, Chunxia Zhang, Cong Ma, Hongtao Wang, Lizhen Ji, Fei Gao, Jiangshe Zhang, and Kaiwen Qiu

Applied Soft Computing · Volume 197 · Article 115168 · 2026

Reinforcement learningAlgorithmic tradingDistributed learningSelf-imitationCSI-DDQN
View published article
Abstract +

In recent years, reinforcement learning has emerged as a prominent approach in algorithmic trading, with most studies concentrating on time-series processing and neural-network architecture optimization within single-agent frameworks. Some research has confirmed that multi-agent or ensemble learning can enhance performance, but conventional methods often train ensemble neural networks or diverse reinforcement learning algorithms on the entire time-series dataset. These practices not only lead to high computational costs but also hinder the efficient recognition of market fluctuations.

To address these challenges, this paper introduces a novel Collaborative Self-Imitation Double Deep Q Network (CSI-DDQN) framework. The proposed approach partitions long-term time-series data into shorter segments and employs a distributed-learning-inspired collaborative training strategy across multiple clients. A self-imitation strategy enhances cooperation among clients, while a hybrid multi-loss function integrates the strengths of reinforcement learning and imitation learning.

Experimental results using a multi-layer perceptron architecture on U.S. stock data demonstrate that the proposed method significantly outperforms several baselines, achieving the highest average cumulative return of 56.85%, the highest Sharpe ratio of 1.42, and the lowest maximum drawdown of 17.14%. In comparison, DQN-Vanilla attains an average cumulative return of 3.95% and a Sharpe ratio of 0.25, while the DRL-Ensemble method yields a cumulative return of 19.53% and a Sharpe ratio of 0.54. Additional experiments substituting various network architectures and reinforcement-learning algorithms further validate the flexibility and effectiveness of the CSI-DDQN framework.

For a complete academic record, teaching history, and references:

Download curriculum vitae