Learning to Generate Explainable Stock Predictions using Self-Reflective Large Language Models (2024)

1. Introduction

The Efficient Market Hypothesis (EMH) states that in financial markets, stock prices reflect all available information (Fama, 1970), and should only react to new information. Through mining and analysing external data sources, the goal of investors is to quickly understand the impact of new information on the market, in order to anticipate future stock price movements (Gidofalvi and Elkan, 2001). However, analyzing the impact of these data on the stock market is a huge undertaking and imposes a heavy workload on financial experts, due to the large volume of information available (Feng etal., 2021). Because of this, many have explored the use of deep-learning techniques (Feng etal., 2019; Li etal., 2021; Koa etal., 2023) for stock prediction.

However, due to their complex and quantitative nature, traditional deep-learning methods in stock prediction areblack box models and do not address the explainability of their predictions (Li etal., 2023a). This reduces their usability in practical applications, as users might notbe able totrust (Biran and McKeown, 2017) the results to invest their capital. Even amongworks that deal with explainable stock predictions (Carta etal., 2021; Li etal., 2023a), the ”explanations” are oftensimplydefined as the specific texts that caused the price movement, which are usually obtained by analyzing learnt attention weights (Deng etal., 2019; Sawhney etal., 2020). For example, these models could analyze a series of texts regarding Apple stock and determine that its Positive prediction is attributed tothe text”Apple reportedrevenue of $90.1 billion, beating expectations”.However, these models do not go beyond that to explain why these texts caused the stock movement, and require the user to make their own inference.

Today, the emergence of Large Language Models (LLMs) has presented a solution to this problem. Recent surveys (Yang etal., 2023a; Zhao etal., 2023) have shown that LLMs possess both strong Natural-Language Understanding capabilities, which allow them to perform tasks like text summarization (Pu etal., 2023) and text classification (Liang etal., 2022) in a few-shot manner; and strong Natural-Language Generation capabilities, which let themgenerate human-readable explanations for their own decision-making process (Liu etal., 2023; Wei etal., 2022). Currently, works that utilize LLMs for stock prediction (Yu etal., 2023; Chen etal., 2023) are few, and use limited techniques such as pre-trained LLMs or instruction tuning. Our work seeks to fill this gap by designing a reinforcement learning (RL) framework which can fine-tune a LLM to generateexplanations forstock prediction.

To tackle the explainable stock prediction task using LLMs, we can identify two main challenges. Firstly, it is well-established in past stock prediction literature that social texts are chaotic, where the influence of different texts on stock prices can be highly diverse (Hu etal., 2018; Xu and Cohen, 2018). For example, breaking news such as surprise earningsannouncements or crisis events often have a visible impact on the stock price, while unsubstantiated opinions or vague remarks usually cause little to no change (Sprenger and Welpe, 2011). This requires a prediction model to have the ability to weigh the varying impacts of new information (Fama and French, 2015), and arrive at a maximum-likelihood prediction (Giglio etal., 2022). Typically, this involves training a regression-based neural network, and is not a known capability of LLMs (see Figure 1). Secondly, the problem becomes progressively harder with the introduction of the explanation component, as it requires the LLM to explain verbally why certain information are more important than others. However, to train a LLM for this task using RL (Ouyang etal., 2022; Hu etal., 2021), one would needgood and bad samples (Lee etal., 2023; Liu etal., 2023) of explanations for each price movement in the training set. This requires substantial amount of labour by financial experts, which is expensive and impractical to scale.

To deal with the above-mentioned problems, we propose our Summarize-Explain-Predict (SEP) framework, which utilizes a self-reflective agent (Shinn etal., 2023) and Proximal Policy Optimization (PPO) (Schulman etal., 2017) to let a LLM teach itself how to make explainable stock predictions in a fully autonomous manner (see Figure 2). Firstly, the Summarize module utilizes the strong summarization capabilities of LLMs (Pu etal., 2023) to convert large volumes of text input data into point-form summaries of factual information. Secondly, in the Explain module, a reflective agent teaches itself to generate correct stock predictions and explain their reasoning (Wei etal., 2022) given a sequence of summarized facts, via an iterative, verbal self-reflective process (Madaan etal., 2023; Shinn etal., 2023). The iterative process additionally allows us to obtain a series of correct and incorrect predictions with annotated explanations through its past mistakes, which can be used as fine-tuning samples without human-in-the-loop. Lastly, in the Predict module, a specialized LLM is fine-tuned (Ouyang etal., 2022; Hu etal., 2021) via PPO training (Schulman etal., 2017) using its own self-taught responses, in order to generate the most likely stock predictions and explanations, given the input texts from an unseen test set.

To demonstrate the effectiveness of the SEP framework, we validate through experimental results that our model is able to outperform both traditional deep-learning and LLM methods in terms of its prediction accuracy and Matthews correlation coefficient (MCC) for the binary stock classification task. We also analyze some responses from the fine-tuned LLM qualitatively, to show how it is better able to understand and weigh the impacts of different information within the input texts. Additionally, to justify the generalization capability of the framework, we testit on the portfolio construction task, by generating explainable weights for a stock portfolio. We also demonstrate the effectiveness of this method through portfolio metrics, such as its profitability and Sharpe Ratio.

2. Related Works

In this section, we trace the progress of textual analysis techniques in stock prediction works,and also explore some pioneering works that utilized Large Language Models (LLMs) in the financial domain.

Text Analysis in Stock Prediction.Early text analysis works in stock prediction first studied the effectiveness of using different textual representations of news, such as Bag of Words, Noun Phrases, and Named Entities, in Support Vector Machines (SVM) (Schumaker and Chen, 2009). These ”shallow” features were later replaced in favor of structured information, where events in the form of (Actor, Action, Object) tuples were used as inputs for deep neural networks (Ding etal., 2014, 2015).

Later works would define the challenges in text analysis more clearly, which was attributed to the chaotic and diverse nature of text data (Hu etal., 2018). This led to the popular use of attention-based models to capture the ”most important” information in texts directly from pre-trained text embeddings (Deng etal., 2019). Some other notable works include the use of Variational Auto-Encoders (VAEs) to model the latent factors in market information (Xu and Cohen, 2018), and Transformer models (Yang etal., 2020).

Most recent works have moved away from improving text analysis methods, and opted instead to enhance the current models with additional forms of information, such as the vocal features from audio data (Yang etal., 2022) or cross-stockimpacts from company relational graphs (Sawhney etal., 2020; Li etal., 2021). In contrast, our work return to purely text-based models, to isolate the effects of text information on stock movements.

Large Language Models in Finance.Out of the existing works that utilize LLMs on general financial tasks, the most well-known one is BloombergGPT (Wu etal., 2023b),which trained a 50B parameters LLM using their existing large financial text corpus. Their model was evaluated on several downstream tasks such as sentiment analysis and named-entity recognition (NER), with optimistic results. Along this direction, some works have also attempted to fine-tune their own financial LLM, which include FinMA (Xie etal., 2023) and FinGPT (Yang etal., 2023b).

Other works explored the use of existing LLMs such as ChatGPT to perform specialized downstream tasks, such as stock sentiment prediction from news headlines (Lopez-Lira and Tang, 2023), and classification of Federal announcements (Hansen and Kazinnik, 2023). These early works focused on analyzing individual texts, as opposed to a sequence of texts. More recent works have explored the use of LLMs to make stock predictions using sequences of stock-related texts, via instruction-tuning (Yu etal., 2023) or pre-trained models enhanced with relational graphs (Chen etal., 2023). We build on these works by implementing an additional verbal self-reflective agent to learn how to generate better explanations, and a PPO trainer to fine-tune a more specialized LLM for stock predictions.

3. Methodology

In this section, we first define the task and data for explainable stock prediction. We then present the proposed SEP framework, which was illustrated in Figure 2. There are three main components: (1) a Summarize module, which generates a summary of factual information from the unstructured text inputs; (2) an Explain module, which generates explanations for its stock predictions and refines them through an iterative self-reflective process; and (3) a Predict module, which generates confidence-based predictions after fine-tuning a LLM using its self-generated annotated samples.

3.3. Explanation Generation

The goal of the Explain module is two-fold: While the key aim of the module is to generate clear explanations for stock predictions, the generated explanations also serve as a reasoning step (Wei etal., 2022) for the LLM to do self-reflection to improve its own predictions (Shinn etal., 2023). In the following subsections, we discuss the initial prompt design and the subsequent self-reflective process for the module.

3.3.1. Explanation Prompting

The prompt for the Explain module contains two variable inputs: the specified stock $s$, and a sequence of extracted information that was generated from the previous module.Given these inputs, the LLM $\text{M}_{E}$ then generate the response $\mathbf{\mathcal{Y}}^{s}_{t}$, which should contain the next-day price movement $\mathbf{y}_{t}^{s}$, and the annotated explanation $\mathbf{e}_{t}^{s}$, i.e., $\mathbf{\mathcal{Y}}^{s}_{t}=(\mathbf{y}_{t}^{s},\mathbf{e}_{t}^{s})$. We formalize this as:

(2)

$$\vspace{-2px}\mathbf{\mathcal{Y}}^{s}_{t}=\text{M}_{E}\left(s,\mathbf{X}_{t-T}$$

Similar to the previous summarization prompt, we select two cases from the dataset and manually compose the response trajectories to use as few-shot exemplars (Ye etal., 2023). Additionally, the two example cases chosen have specifically one Positive and one Negative movement label, in order to avoid any majority label bias (Zhao etal., 2021). The prompt trajectories are designed in a fashion similar to ReAct (Yao etal., 2022), albeit in a singular, prediction-explanation step.

3.3.2. Self-Reflective Process

Current LLMs are not trained to generate stock predictions, which could cause incorrectly-generated annotated examples in the previous step. To tackle this, we deploy the LLM as an autonomous agent that can iteratively improve on its past responses, through a verbal self-reflection loop (see Figure 3).The loop is first seeded with the response from the previous step, i.e., $\mathbf{\mathcal{Y}}^{s}_{t,0}=\mathbf{\mathcal{Y}}^{s}_{t}$, which is taken to be the initial iteration $i=0$.

From the generated price movement $\mathbf{y}_{t,i}^{s}$, we can obtain a binary feedback by evaluating its alignment with the ground truth. For the incorrect samples, we then prompt a LLM $\text{M}_{R}$ to generate a verbal feedback $\mathbf{r}_{t,i}^{s}$ for each iteration $i$, given its previous inputs and outputs, which we refer to as its short-term memory (Shinn etal., 2023). The feedback should explain clearly where it went wrong in its previous reasoning $\mathbf{e}_{t,i}^{s}$, and also come up with a high-level plan to mitigate this failure for the next iteration. The overall formalization is:

(3)

$$\vspace{-1px}\mathbf{r}_{t,i}^{s}=\text{M}_{R}\left(s,\mathbf{X}_{t-T}^{s},$$

For every iteration, each reflection $\mathbf{r}_{t,i}^{s}$ represent a lesson that the LLM learnt from its failures, which is added to its experiences, or long-term memory (Shinn etal., 2023). We represent this as a set of reflections,$\mathbf{R}_{t,i}^{s}=\left[\mathbf{r}_{t,0}^{s},\mathbf{r}_{t,1}^{s},\cdots,$.The reflections, together with the original inputs, are fed again into LLM $\text{M}_{E}$ to generate the price movement and explanation for the next iteration. The formalization is:

(4)

$$\vspace{-1px}\mathbf{\mathcal{Y}}^{s}_{t,i}=\text{M}_{E}\left(s,\mathbf{X}_{t-$$

The prompt and response examples can be found in Appendix B.

Through this process, we are then able to obtain pairs of correct and incorrect responses, for each successful reflection. We define these as $\mathbf{\mathcal{Y}}_{w,t}^{s}=\left(\mathbf{y}_{t,\tilde{i}}^{s},\mathbf{e}_{$ and $\mathbf{\mathcal{Y}}_{l,t}^{s}=\left(\mathbf{y}_{t,\tilde{i}-1}^{s},\mathbf{e}$ respectively, where $\tilde{i}$ refers to the iteration in which the reflective process resulted in the LLM $\text{M}_{E}$ generating the correct stock movement.

3.4. Prediction Generation

The goal of the Predict module is to fine-tune a LLM to generate good stock predictions and explanations for the unseen test period. In this section, we discuss the overall fine-tuning process of the model and the subsequent inference procedure at test-time.

3.4.1. Model Fine-Tuning

Following previous works that tackles Reinforcement Learning from Human Feedback (RLHF) (Ouyang etal., 2022; Stiennon etal., 2020), we utilize a similar three-step process to fine-tune a LLM. Instead of human feedback, we use the binary evaluations from the reflections to choose the ”better” response during training (see Figure 4).

In the first step, we collect the demonstration data, which are taken from the correct predictions in the initial iteration $\mathbf{\mathcal{Y}}^{s}_{t,0}$. These samples do not have corresponding ”wrong” responses, as they were taken from the initial prompt. The samples are used to train a supervised policy $\pi^{SFT}$ using Supervised Fine-Tuning (SFT).

In the second step, we collect the comparison data $\mathcal{D}$, which contains pairwise correct and incorrect responses $\mathbf{\mathcal{Y}}_{w,t}^{s},\mathbf{\mathcal{Y}}_{l,t}^{s}$ for each structured input $\mathbf{X}^{s}_{t}$, taken from the successful reflection iterations. These are used to train a reward model $r_{\theta}$, which learns to give higher reward scores to the correct responses. Specifically, we train the model to minimize the following cross-entropy loss (Stiennon etal., 2020):

(5)

$$\vspace{-1px}\mathcal{L}(\boldsymbol{\theta})=-\mathbb{E}_{\left(\mathbf{X},$$

See Also

A Survey of Large Language Models for Financial Applications: Progress, Prospects and ChallengesRevolutionizing Finance with LLMs: An Overview of Applications and InsightsWhat are Large Language Models? | A Comprehensive LLMs GuideStock Market Theories — Yochaa

In the third step, we use the reward model to optimize the trained policy using PPO (Schulman etal., 2017). We first initialize the model with the supervised policy $\pi^{SFT}$, and use it to generate predictions $\mathbf{\hat{\mathcal{Y}}}^{s}_{t}$ for randomly selected samples $\mathbf{X}^{s}_{t}$ from the overall dataset $\mathcal{D}_{\pi_{\phi}^{RL}}$. Next, the reward model $r_{\theta}$ is used to generate a reward for each response. We then try to optimize a PPO model $\pi_{\phi}^{RL}$ by maximizing the overall reward. This is achieved by minimizing the following loss objective:

(6)

$$\vspace{-2px}\mathcal{L}(\boldsymbol{\phi})=-\mathbb{E}_{\left(\mathbf{X},$$

We note that the objective includes an additional term that penalizes the KL divergence between the trained policy $\pi_{\phi}^{RL}$ and the supervised policy $\pi^{SFT}$ (Jaques etal., 2019), which is used to deter the policy from collapsing into a single mode (Stiennon etal., 2020), and prevent it from generating responses that are too different from those of the original reference model $\pi^{SFT}$ (Yao etal., 2023). The term is controlled by a hyper-parameter $\beta$.

3.4.2. Confidence-based Sampling

During inference, the unstructured input texts $\mathbf{\mathcal{C}}_{t}^{s}$ are first summarized using a pre-trained LLM. We then use the trained policy $\pi_{\phi}^{RL}$ to generate the next-day predictions $\mathbf{\hat{\mathcal{Y}}^{s}}_{t}$ from the summarized facts $\mathbf{X}_{t}^{s}$. For generating predictions, we use a best-of-$n$ sampler, where we generate $n$ responses and use the scores from reward model $r_{\theta}$ to select the best response (Yao etal., 2023).

4. Experiment

4.2. Performance Comparison (RQ1)

In this section, we evaluate both the prediction and explanation responses generated by our SEP model, through quantitative and qualitative comparisons against the relevant baselines.

4.2.1. Prediction Performance

Table 1 reports the quantitative results on the stock prediction task. On the prediction accuracy, we observe that the SEP model fine-tuned on the GPT-generated explanations (Table 1, left) was able to obtain the best results, achieving an improvement of 2.4% over the strongest baseline (GRU$+$Att) using all texts. On the other hand, the SEP model fine-tuned on explanations generated by Vicuna-v1.5 (Table 1, right) under-performed the baselines in terms of accuracy. A possible reason for this is that the Vicuna-generated explanations used for training the model are prone to hallucinations, which could negatively impact the reasoning ability of the SEP model (see Figure 5).The poorer performance of GPT-3.5, a pre-trained LLM, is largely attributed to its inability to make decisive predictions from mixed sentiments. The instruction-tuned FinGPT-Forecaster is able to improve on this by guiding the LLM towards trained responses, which are in the correct format. Finally, our SEP model produces the best accuracy, likely due to its additional self-reflective process and reinforcement learning.

Models		Top 1 Stock, GPT-3.5				Remaining Stocks, Vicuna
		All Texts		Informative Texts		All Texts		Informative Texts
		Accuracy	MCC	Accuracy	MCC	Accuracy	MCC	Accuracy	MCC
Deep-LearningModels	VAE+Att	49.96	0.0046	-	-	49.83	0.0070	-	-
	GRU+Att	50.15	0.0125	-	-	50.77	0.0189	-	-
	Transformer	50.06	0.0089	-	-	50.17	0.0135	-	-
Large LanguageModels	GPT-3.5	20.80	0.0094	29.35	0.0298	17.57	0.0027	22.99	0.0052
	Vicuna	40.85	0.0114	45.29	0.0368	39.66	0.0115	43.30	0.0301
	FinGPT	47.61	0.0158	51.56	0.0384	45.76	0.0161	46.12	0.0379
	SEP (Ours)	51.38	0.0302	54.35	0.0993	47.59	0.0203	50.57	0.0508

For this task, a more telling metric is the Matthews Correlation Coefficient (MCC), which takes into account the ratios of True and False Positives and Negatives of the predictions (Chicco and Jurman, 2020; Chicco etal., 2021). Given that not all stock movements are necessarily caused by the provided texts, the accuracy results might not be fully indicative of the model’s natural language processing capabilities, as it includes some random guesses on the non-informative texts. After filtering for informative texts only, we can see increases in the MCC ratio,possibly from havingless random guesses in the prediction results.

On the MCC metric, our SEP model was able to outperform all models under all settings, which showcase the true ability of the model to understand the impacts of natural language texts on stock movements, after accounting for the random guesses. Under the all-texts setting, we are able to outperform the strongest deep-learning baseline (GRU$+$Att) by 0.0177 for the GPT-3.5-based model, and 0.0014 for the Vicuna-based model.After filtering for informative texts only, our fine-tuned SEP model is also able to outperform the strongest LLM baseline, FinGPT-Forecaster, by 0.0609 and 0.0129 for the GPT-3.5 and Vicuna-based SEP models respectively.

4.2.2. Explanation Performance

In addition to generating better predictions, the natural advantage of using LLMs over traditional deep-learning methods is their capability to generate explanations for their predictions. Here, we compare the generated explanations qualitatively between the pre-trained LLMs and our SEP model.

After SEP fine-tuning, we can observe two main improvements. The first deals with the ability to decisively weigh between news information to make a stock movement prediction. While pre-trained LLMs are known to be able to classify the sentiment of individual texts (Zhang etal., 2023; Lopez-Lira and Tang, 2023), they typically do not try to weigh between these sentiments and make a decisive stock prediction, even if specifically requested by the prompt (see Figure 1). This is generally an easier task to tackle, which is similar to fine-tuning an expert LLM (Guo etal., 2023), albeit ours is trained without human experts-in-the-loop. Figure 6 shows an example of how our SEP model can learn how to make a decisive stock prediction after the fine-tuning process.

The second improvement deals with the ability to generate better-quality explanations. This is a more difficult task for LLMs, as it requires them to not only understand the meaning of natural language texts, but also to correctly reason out their overall impact on the stock price movement. Through the SEP framework, our LLM first learns to reason out the correct explanations via self-reflection and teach them to the PPO model, which learns to determine heuristically what is the most probable explanation at test-time. For this comparison, we came up with a set of metrics for explanation quality, and use GPT-4 to rate each response from 1 to 7 for the samples in the top-1 stocks. The average score for each metric is reported in Table 2. Explanations of the metrics can be found in Appendix C.

From Table 2, we can make the following observations:

• •

  The highest scores come from more generalizable metrics, such as Consistency with Information. Some metrics are sometimes not observable given that no such information is in the texts, which will lower their scores. However, it is fair to compare the relative scores across the LLMs, as their input texts are the same.

• •

  All LLMs give good-quality explanations even if the prediction is wrong, as they can naturally understand the input texts and make reasonable comparisons, but have done so incorrectly. Thus, prediction accuracy should still be the first metric to look at.

• •

  Our SEP model was able to score the highest for all metrics. We note that the model was not trained on these metrics, but was only provided a reward based on correct binary predictions. Through its self-reflection and reinforcement learning, the SEP framework was able to intrinsically teach the model to better compare these factors, in order to generate better predictions.

Metric	GPT-3.5	Vicuna	SEP (Ours)
Relevance to Stock Movement	5.407	5.396	5.449
Financial Metrics	2.957	3.146	3.334
Global & Industry Factors	3.180	3.576	3.700
Company Developments	3.905	4.066	4.224
Temporal Awareness	3.951	4.066	4.170
Balance of Positive & Negative	4.030	4.084	4.224
Contextual Understanding	4.012	4.098	4.193
Clarity & Coherence	6.271	6.325	6.439
Consistency with Information	5.575	5.652	6.006
Sensitivity to Updates	4.112	4.172	4.362

4.3. Ablation Study (RQ2)

In this section, we evaluate the efficiency of each of the three proposed components: the Summarize, Explain and Predict modules.

4.3.1. Summarize Module

The Summarize module reduces the noise and length of the input texts by extracting only the important, factual information. For the ablation study, we compare against the performance of using non-summarized, raw social texts in our trained SEP model. To keep the input lengths within the LLM’s token limit, we try two ways of using the raw texts: Taking the 30 most shared texts and randomly sampling 30 texts, for each day.

	Top 1 Stock (GPT-3.5)		Remaining Stocks (Vicuna)
	Accuracy	MCC	Accuracy	MCC
Non-Summ. (Random 30)	50.75	0.0208	40.81	-0.0037
Non-Summ. (Top 30)	50.81	0.0219	41.27	0.0023
Summarized (All Texts)	51.38	0.0302	47.59	0.0203
Summarized (Informative)	54.35	0.0993	50.57	0.0508

From Table 3, we can make the following observations:

• •

  Using the most shared texts is better than randomly sampling.

• •

  The model trained on Vicuna-generated responses fare much worse without summarizing than the GPT-3.5-trained model. This could be attributed to the texts causing more hallucination (Table 5), given their chaotic content (e.g., emojis, spam, etc. ).

• •

  The summarized texts provide better results. One possible reason here could be due to having information from more than 30 texts. However, it would also show that the summarization process did not lose any important information that would cause degradation.

  See Also

  NLP in Finance: Examining the Impact of Natural Language Processing in Financial and Banking Services

• •

  Finally, removing the non-informative texts, which is only possible with the Summarize module, provides the best results.

4.3.2. Explain Module

In the SEP model, we have observed two improvements: 1) the ability to make decisive stock predictions from mixed sentiments; and 2) the ability to make correct stock predictions (i.e., better prediction accuracy). In order to fine-tune the LLM to produce these predictions and explanations, the Explain module must first try to generate correctly-annotated samples through binary feedback and self-reflection. To demonstrate its effectiveness, we plot the percentage change in the number of generated decisive and correct predictions after each of its reflective iteration.

From Figure 7, we see that with multiple self-reflective iterations, the model generates more and more decisive and correct annotated samples, to be used for fine-tuning.We also observe that there is a greater number of decisive samples produced given that it is an easier task, which starts to slow down as more samples become non-Neutral. Overall, the number of decisive samples grew by 49.8% while the correct samples grew by 43.2% after 3 iterations, which highlights the effectiveness of the Explain module in generating annotated explanation samples, without the help of human experts.

4.3.3. Predict Module

For the Predict module, we conduct an ablation study over different variants of the model. We remove one additional component for each variant, i.e., no $n$-shot sampling at inference [SEP (1-shot)]; no PPO reinforcement learning [SEP (no PPO)]; and no explanations [SEP (binary)], which is simply instruction-tuning the LLM to make binary up/down predictions. We make the comparisons on the top-1 stock from each industry.

	All Texts		Informative Texts
	Accuracy	MCC	Accuracy	MCC
SEP (binary)	40.84	-0.0042	42.75	0.0295
SEP (no PPO)	44.08	0.0144	45.29	0.0368
SEP (1-shot)	50.08	0.0270	52.54	0.0715
SEP (Ours)	51.38	0.0302	54.35	0.0993

From Table 4, we can make the following observations:

• •

  The addition of the explanation component during the instruction-tuning process, i.e., from SEP (binary) to SEP (no PPO), gives the model an average improvement of 6.9%. It is likely that by tuning the LLM to generate explanations, we are able to elicit a reasoning process from the LLM (Wei etal., 2022) when generating stock movement predictions, resulting in better prediction accuracy.

• •

  The instruction-tuned variant, i.e., SEP (no PPO), shows very similar results to the base model that it is tuned on (i.e., the Vicuna model in Table 1). It is possible that the instruction tuning process has no impact on the SEP model given that the samples, taken before the reflective iterations (i.e., Step 1 in Figure 4), are ”easy” samples that the base model could already handle. We also note that supervised-tuned models have been seen to produce little to even negative improvements in previous literature (Stiennon etal., 2020).

• •

  The largest improvement comes from the PPO reinforcement learning, i.e., from SEP (no PPO) to SEP (1-shot), with an average improvement of 14.8%. This highlights the ability of the PPO trainer in teaching the LLM to generate stock predictions more effectively. Additionally, the $n$-shot sampling weighs between $n$ generated samples using the learnt reward model to select the best output. The average improvement of this variant i.e., 3.0% from SEP (1-shot) to SEP (Ours), further reinforces the usefulness of the reward model trained during the PPO process.

4.4. Portfolio Optimization (RQ3)

From our results, we have observed that the SEP framework is able to teach an LLM to weigh the impact of information within the input texts in a binary manner. We further explore its generalization capability by using it to fine-tune a LLM to weigh between information within its own generated explanations quantitatively, in order to generate portfolio weights for the stock portfolio task.

For the portfolio task, we follow the same method as above to fine-tune a LLM. Here, the input information are now all the generated explanations for the basket of stocks for each day. For this experimental task, we filter only the stocks with positive predictions, in order to reduce the number of stocks the LLM have to weigh, and to prevent negative weights (hence setting a no short-sales constraint (Koa etal., 2023)). We then prompt the LLM to generate portfolio weights given the outlook for each given stock (see Figure 8). A full example of the prompt and response can be found in Appendix B.

As there is no binary feedback for this task, in each self-reflective iteration, we provide the reflective LLM with the overall profits based on the provided portfolio weights, and prompt it to reflect on how it can improve itself to obtain higher profits. The reflections are then used to generate an updated set of portfolio weights. Finally, we feed both sets of generated weights into a PPO trainer, where the one with higher profits is used as the ”better” response.

We compare the performances of portfolios generated by three different LLMs: GPT-3.5-turbo, Vicuna, and our fine-tuned SEP model. We also include three baselines: the 1/N portfolio, where all 11 stocks in the basket are bought at equal weights (DeMiguel etal., 2009); the S&P500 stock market index; and Positive-Only, where only the predicted positive stocks are bought at equal weights. The latter can also be seen as evaluating the results of the original stock prediction LLM in a practical setting, without the portfolio weighing prompts.

We evaluate the portfolio performance using four metrics: the overall gain, which simply sums up the gains for each day; the cumulative gain, which is the final gain after re-investing any additional profits or losses over the evaluation period; the standard deviation of the profits; and the annualized Sharpe Ratio (Lo, 2002).

Approach	Overall	Cumulative	Std. Dev.	Sharpe
1/N	-0.0330	-0.0502	1.613e-2	-0.225
Market Index	0.0180	0.0003	1.533e-2	0.123
Positive-Only	0.1243	0.1065	1.911e-2	0.807
GPT-3.5	0.1497	0.1353	1.893e-2	0.980
Vicuna	0.1541	0.1447	1.731e-2	1.104
SEP (Ours)	0.1661	0.1569	1.792e-2	1.150

Table 5 reports the portfolio results. From the table, we observe:

• •

  The Positive-Only portfolio, i.e., evenly buying the stocks that are predicted to be Positive, already showcases good performance. This highlights the capability of our original stock prediction model to produce good trading signals in a practical setting.

• •

  For the standard deviation results, we note that the top 2 portfolio methods, i.e., 1/N and Market Index, contains more number of stocks, which allow them to spread out the stock price fluctuations more evenly. However, their Sharpe Ratios are still lower than the other models, which shows a lower reward-to-risk ratio.

• •

  The pre-trained LLM models, i.e., GPT-3.5 and Vicuna, already shows better performance than the Positive-Only portfolio in most metrics, which shows the capabilities of using LLMs to weigh between information to produce portfolio weights.

• •

  Our SEP model was able to outperform all other methods in most portfolio metrics,and achieve comparable performance in its standard deviation,which showcases the effectiveness of our SEP framework.In addition to the shown metrics, we also re-emphasize the ability of the LLM-based models to explain the generated portfolio weights, which further adds to the interpretability and trustability of their results for practitioners.

5. Conclusion and Future Work

In this work, we explored the explainable stock prediction task, which was largely difficult to solve before generative models. For this task, we highlighted two challenges: the limitations of current LLMs in weighing varied market factors to make aggregate stock predictions, and the lack of annotated training samples for fine-tuning LLMs to make explanations. To tackle these challenges, we proposed our SEP framework, which utilizes a verbal self-reflective agent and PPO techniques to let a LLM teach itself how to generate stock explanations in a fully autonomous manner. Through experimental results, we validated that our SEP model is able to outperform both traditional deep-learning and LLM methods in the accuracy of the predictions and quality of the generated explanations. Furthermore, we also demonstrated the generalizability of the SEP framework by fine-tuning a model for the portfolio task.

There are some directions that can be explored in future works. Firstly, we address the possibility of cumulative errors in the SEP framework. At each stage, poorly generated summaries or explanations could lead to poorer responses in the next step. In practice, it is possible for experts to vet through the responses before using them, which would be an easier task than generating them manually. However, more can be done to increase the robustness of the generated responses and reduce the need for human-in-the-loop. Secondly, using additional data sources, such as knowledge graphs (Hansen and Kazinnik, 2023) or audio features (Yang etal., 2022), could increase the quality of the predictions. At the same time, such works would also help to explore the multi-modal capabilities of the most recent LLM upgrades (Bang etal., 2023; Wu etal., 2023a).Finally, as this is a relatively new task, there are currently limited works on evaluating the generated stock explanations. Further studies can be done to improve the metrics created in this work.

6. Ethical Use of Data

For this research, we have utilized datasets derived from publicly available sources, and no human annotators were involved in the data collection process. Rights pertaining to the data used, such as text data, remain the sole property of the original rights holders.This study is intended exclusively for academic purposes only.

In general, the most effective mitigation strategy is to include human-in-the-loop to anticipate and mitigate various potential risks. While LLMs can assist humans in labor-intensive tasks such as processing large volume of texts and analyzing their stock market impacts, it is not able to replace the need for human oversight.

7. Acknowledgement

This research is supported by the National Research Foundation, Singapore under its Industry Alignment Fund – Pre-Positioning (IAF-PP) Funding Initiative, by the National Research Foundation, Singapore through the National Cybersecurity R&D Lab at the National University of Singapore under its National Cybersecurity R&D Programme (Award No. NCR25-NCL P3-0001). Any opinions, findings and conclusions or recommendations expressed in this material are those of the author(s) and do not reflect the views of National Research Foundation, Singapore.

References

• (1)
• Alam and Uddin (2009)MdMahmudul Alam and Gazi Uddin. 2009.Relationship between interest rate and stock price: empirical evidence from developed and developing countries.International Journal of Business and Management (ISSN 1833-3850) 4, 3 (2009), 43–51.
• Bahdanau etal. (2014)Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. 2014.Neural machine translation by jointly learning to align and translate.arXiv preprint arXiv:1409.0473 (2014).
• Bang etal. (2023)Yejin Bang, Samuel Cahyawijaya, Nayeon Lee, Wenliang Dai, Dan Su, Bryan Wilie, Holy Lovenia, Ziwei Ji, Tiezheng Yu, Willy Chung, etal. 2023.A multitask, multilingual, multimodal evaluation of chatgpt on reasoning, hallucination, and interactivity.arXiv preprint arXiv:2302.04023 (2023).
• Biran and McKeown (2017)Or Biran and KathleenR McKeown. 2017.Human-Centric Justification of Machine Learning Predictions.. In IJCAI, Vol.2017. 1461–1467.
• Campello etal. (2013)RicardoJGB Campello, Davoud Moulavi, and Jörg Sander. 2013.Density-based clustering based on hierarchical density estimates. In Pacific-Asia conference on knowledge discovery and data mining. Springer, 160–172.
• Carta etal. (2021)SalvatoreM Carta, Sergio Consoli, Luca Piras, AlessandroSebastian Podda, and DiegoReforgiato Recupero. 2021.Explainable machine learning exploiting news and domain-specific lexicon for stock market forecasting.IEEE Access 9 (2021), 30193–30205.
• Casper etal. (2023)Stephen Casper, Jason Lin, Joe Kwon, Gatlen Culp, and Dylan Hadfield-Menell. 2023.Explore, Establish, Exploit: Red Teaming Language Models from Scratch.arXiv preprint arXiv:2306.09442 (2023).
• Cer etal. (2018)Daniel Cer, Yinfei Yang, Sheng-yi Kong, Nan Hua, Nicole Limtiaco, RhomniSt John, Noah Constant, Mario Guajardo-Cespedes, Steve Yuan, Chris Tar, etal. 2018.Universal sentence encoder.arXiv preprint arXiv:1803.11175 (2018).
• Chen and Shu (2023)Canyu Chen and Kai Shu. 2023.Combating misinformation in the age of llms: Opportunities and challenges.arXiv preprint arXiv:2311.05656 (2023).
• Chen etal. (2023)Zihan Chen, LeiNico Zheng, Cheng Lu, Jialu Yuan, and Di Zhu. 2023.ChatGPT Informed Graph Neural Network for Stock Movement Prediction.arXiv preprint arXiv:2306.03763 (2023).
• Chiang etal. (2023)Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhanghao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yonghao Zhuang, JosephE Gonzalez, etal. 2023.Vicuna: An open-source chatbot impressing gpt-4 with 90%* chatgpt quality.See https://vicuna. lmsys. org (accessed 14 April 2023) (2023).
• Chicco and Jurman (2020)Davide Chicco and Giuseppe Jurman. 2020.The advantages of the Matthews correlation coefficient (MCC) over F1 score and accuracy in binary classification evaluation.BMC genomics 21, 1 (2020), 1–13.
• Chicco etal. (2021)Davide Chicco, Niklas Tötsch, and Giuseppe Jurman. 2021.The Matthews correlation coefficient (MCC) is more reliable than balanced accuracy, bookmaker informedness, and markedness in two-class confusion matrix evaluation.BioData mining 14, 1 (2021), 1–22.
• DeMiguel etal. (2009)Victor DeMiguel, Lorenzo Garlappi, and Raman Uppal. 2009.Optimal versus naive diversification: How inefficient is the 1/N portfolio strategy?The review of Financial studies 22, 5 (2009), 1915–1953.
• Deng etal. (2019)Shumin Deng, Ningyu Zhang, Wen Zhang, Jiaoyan Chen, JeffZ Pan, and Huajun Chen. 2019.Knowledge-driven stock trend prediction and explanation via temporal convolutional network. In Companion Proceedings of The 2019 World Wide Web Conference. 678–685.
• Devlin etal. (2018)Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2018.Bert: Pre-training of deep bidirectional transformers for language understanding.arXiv preprint arXiv:1810.04805 (2018).
• Ding etal. (2014)Xiao Ding, Yue Zhang, Ting Liu, and Junwen Duan. 2014.Using structured events to predict stock price movement: An empirical investigation. In Proceedings of the 2014 conference on empirical methods in natural language processing (EMNLP). 1415–1425.
• Ding etal. (2015)Xiao Ding, Yue Zhang, Ting Liu, and Junwen Duan. 2015.Deep learning for event-driven stock prediction. In Twenty-fourth international joint conference on artificial intelligence.
• Ding etal. (2021)Yujuan Ding, Yunshan Ma, Lizi Liao, WaiKeung Wong, and Tat-Seng Chua. 2021.Leveraging multiple relations for fashion trend forecasting based on social media.IEEE Transactions on Multimedia 24 (2021), 2287–2299.
• Fama (1970)EugeneF Fama. 1970.Efficient capital markets: A review of theory and empirical work.The journal of Finance 25, 2 (1970), 383–417.
• Fama and French (2015)EugeneF Fama and KennethR French. 2015.A five-factor asset pricing model.Journal of financial economics 116, 1 (2015), 1–22.
• Feng etal. (2018)Fuli Feng, Huimin Chen, Xiangnan He, Ji Ding, Maosong Sun, and Tat-Seng Chua. 2018.Enhancing stock movement prediction with adversarial training.arXiv preprint arXiv:1810.09936 (2018).
• Feng etal. (2019)Fuli Feng, Xiangnan He, Xiang Wang, Cheng Luo, Yiqun Liu, and Tat-Seng Chua. 2019.Temporal relational ranking for stock prediction.ACM Transactions on Information Systems (TOIS) 37, 2 (2019), 1–30.
• Feng etal. (2021)Fuli Feng, Moxin Li, Cheng Luo, Ritchie Ng, and Tat-Seng Chua. 2021.Hybrid learning to rank for financial event ranking. In Proceedings of the 44th International ACM SIGIR Conference on Research and Development in Information Retrieval. 233–243.
• Gallegos etal. (2023)IsabelO Gallegos, RyanA Rossi, Joe Barrow, MdMehrab Tanjim, Sungchul Kim, Franck Dernoncourt, Tong Yu, Ruiyi Zhang, and NesreenK Ahmed. 2023.Bias and fairness in large language models: A survey.arXiv preprint arXiv:2309.00770 (2023).
• Gao etal. (2021)Tianyu Gao, Xingcheng Yao, and Danqi Chen. 2021.Simcse: Simple contrastive learning of sentence embeddings.arXiv preprint arXiv:2104.08821 (2021).
• Gidofalvi and Elkan (2001)Gyozo Gidofalvi and Charles Elkan. 2001.Using news articles to predict stock price movements.Department of computer science and engineering, university of california, san diego 17 (2001).
• Giglio etal. (2022)Stefano Giglio, Bryan Kelly, and Dacheng Xiu. 2022.Factor models, machine learning, and asset pricing.Annual Review of Financial Economics 14 (2022), 337–368.
• Grootendorst (2022)Maarten Grootendorst. 2022.BERTopic: Neural topic modeling with a class-based TF-IDF procedure.arXiv preprint arXiv:2203.05794 (2022).
• Guo etal. (2023)Biyang Guo, Xin Zhang, Ziyuan Wang, Minqi Jiang, Jinran Nie, Yuxuan Ding, Jianwei Yue, and Yupeng Wu. 2023.How close is chatgpt to human experts? comparison corpus, evaluation, and detection.arXiv preprint arXiv:2301.07597 (2023).
• Hansen and Kazinnik (2023)AnneLundgaard Hansen and Sophia Kazinnik. 2023.Can ChatGPT Decipher Fedspeak?Available at SSRN (2023).
• Hu etal. (2021)EdwardJ Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. 2021.Lora: Low-rank adaptation of large language models.arXiv preprint arXiv:2106.09685 (2021).
• Hu etal. (2018)Ziniu Hu, Weiqing Liu, Jiang Bian, Xuanzhe Liu, and Tie-Yan Liu. 2018.Listening to chaotic whispers: A deep learning framework for news-oriented stock trend prediction. In Proceedings of the eleventh ACM international conference on web search and data mining. 261–269.
• Jannati etal. (2023)Sima Jannati, Alok Kumar, Alexandra Niessen-Ruenzi, and Justin Wolfers. 2023.In-group bias in financial markets.Available at SSRN 2884218 (2023).
• Jaques etal. (2019)Natasha Jaques, Asma Ghandeharioun, JudyHanwen Shen, Craig Ferguson, Agata Lapedriza, Noah Jones, Shixiang Gu, and Rosalind Picard. 2019.Way off-policy batch deep reinforcement learning of implicit human preferences in dialog.arXiv preprint arXiv:1907.00456 (2019).
• Kingma and Welling (2013)DiederikP Kingma and Max Welling. 2013.Auto-encoding variational bayes.arXiv preprint arXiv:1312.6114 (2013).
• Koa etal. (2023)KelvinJ.L. Koa, Yunshan Ma, Ritchie Ng, and Tat-Seng Chua. 2023.Diffusion Variational Autoencoder for Tackling Stochasticity in Multi-Step Regression Stock Price Prediction. In Proceedings of the 32nd ACM International Conference on Information & Knowledge Management.
• Lee etal. (2023)Harrison Lee, Samrat Phatale, Hassan Mansoor, Kellie Lu, Thomas Mesnard, Colton Bishop, Victor Carbune, and Abhinav Rastogi. 2023.RLAIF: Scaling Reinforcement Learning from Human Feedback with AI Feedback.arXiv preprint arXiv:2309.00267 (2023).
• Li etal. (2023b)Haochen Li, Maria Polukarov, and Carmine Ventre. 2023b.Detecting financial market manipulation with statistical physics tools. In Proceedings of the Fourth ACM International Conference on AI in Finance. 1–1.
• Li etal. (2023a)Shuqi Li, Weiheng Liao, Yuhan Chen, and Rui Yan. 2023a.PEN: prediction-explanation network to forecast stock price movement with better explainability. In Proceedings of the AAAI Conference on Artificial Intelligence, Vol.37. 5187–5194.
• Li etal. (2021)Wei Li, Ruihan Bao, Keiko Harimoto, Deli Chen, Jingjing Xu, and Qi Su. 2021.Modeling the stock relation with graph network for overnight stock movement prediction. In Proceedings of the twenty-ninth international conference on international joint conferences on artificial intelligence. 4541–4547.
• Liang etal. (2022)Percy Liang, Rishi Bommasani, Tony Lee, Dimitris Tsipras, Dilara Soylu, Michihiro Yasunaga, Yian Zhang, Deepak Narayanan, Yuhuai Wu, Ananya Kumar, etal. 2022.Holistic evaluation of language models.arXiv preprint arXiv:2211.09110 (2022).
• Liu etal. (2023)Hao Liu, Carmelo Sferrazza, and Pieter Abbeel. 2023.Chain of hindsight aligns language models with feedback.arXiv preprint arXiv:2302.02676 3 (2023).
• Liu etal. (2019)Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. 2019.Roberta: A robustly optimized bert pretraining approach.arXiv preprint arXiv:1907.11692 (2019).
• Lo (2002)AndrewW Lo. 2002.The statistics of Sharpe ratios.Financial analysts journal 58, 4 (2002), 36–52.
• Lopez-Lira and Tang (2023)Alejandro Lopez-Lira and Yuehua Tang. 2023.Can chatgpt forecast stock price movements? return predictability and large language models.arXiv preprint arXiv:2304.07619 (2023).
• Ma etal. (2023)Yunshan Ma, Chenchen Ye, Zijian Wu, Xiang Wang, Yixin Cao, Liang Pang, and Tat-Seng Chua. 2023.Structured, Complex and Time-complete Temporal Event Forecasting.arXiv preprint arXiv:2312.01052 (2023).
• Madaan etal. (2023)Aman Madaan, Niket Tandon, Prakhar Gupta, Skyler Hallinan, Luyu Gao, Sarah Wiegreffe, Uri Alon, Nouha Dziri, Shrimai Prabhumoye, Yiming Yang, etal. 2023.Self-refine: Iterative refinement with self-feedback.arXiv preprint arXiv:2303.17651 (2023).
• McInnes etal. (2018)Leland McInnes, John Healy, and James Melville. 2018.Umap: Uniform manifold approximation and projection for dimension reduction.arXiv preprint arXiv:1802.03426 (2018).
• Niessen-Ruenzi and Ruenzi (2019)Alexandra Niessen-Ruenzi and Stefan Ruenzi. 2019.Sex matters: Gender bias in the mutual fund industry.Management Science 65, 7 (2019), 3001–3025.
• Ouyang etal. (2022)Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, etal. 2022.Training language models to follow instructions with human feedback.Advances in Neural Information Processing Systems 35 (2022), 27730–27744.
• Pennington etal. (2014)Jeffrey Pennington, Richard Socher, and ChristopherD Manning. 2014.Glove: Global vectors for word representation. In Proceedings of the 2014 conference on empirical methods in natural language processing (EMNLP). 1532–1543.
• Pu etal. (2023)Xiao Pu, Mingqi Gao, and Xiaojun Wan. 2023.Summarization is (Almost) Dead.arXiv preprint arXiv:2309.09558 (2023).
• Qin etal. (2017)Yao Qin, Dongjin Song, Haifeng Chen, Wei Cheng, Guofei Jiang, and Garrison Cottrell. 2017.A dual-stage attention-based recurrent neural network for time series prediction.arXiv preprint arXiv:1704.02971 (2017).
• Rajaraman and Ullman (2011)Anand Rajaraman and JeffreyDavid Ullman. 2011.Mining of massive datasets.Cambridge University Press.
• Sawhney etal. (2020)Ramit Sawhney, Shivam Agarwal, Arnav Wadhwa, and Rajiv Shah. 2020.Deep attentive learning for stock movement prediction from social media text and company correlations. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP). 8415–8426.
• Schulman etal. (2017)John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. 2017.Proximal policy optimization algorithms.arXiv preprint arXiv:1707.06347 (2017).
• Schumaker and Chen (2009)RobertP Schumaker and Hsinchun Chen. 2009.Textual analysis of stock market prediction using breaking financial news: The AZFin text system.ACM Transactions on Information Systems (TOIS) 27, 2 (2009), 1–19.
• Shinn etal. (2023)Noah Shinn, Federico Cassano, Beck Labash, Ashwin Gopinath, Karthik Narasimhan, and Shunyu Yao. 2023.Reflexion: Language Agents with Verbal Reinforcement Learning.arXiv:2303.11366[cs.AI]
• Sprenger and Welpe (2011)TimmO Sprenger and IsabellM Welpe. 2011.News or noise? The stock market reaction to different types of company-specific news events.The Stock Market Reaction to Different Types of Company-Specific News Events (2011).
• Stiennon etal. (2020)Nisan Stiennon, Long Ouyang, Jeffrey Wu, Daniel Ziegler, Ryan Lowe, Chelsea Voss, Alec Radford, Dario Amodei, and PaulF Christiano. 2020.Learning to summarize with human feedback.Advances in Neural Information Processing Systems 33 (2020), 3008–3021.
• Wei etal. (2023)Alexander Wei, Nika Haghtalab, and Jacob Steinhardt. 2023.Jailbroken: How does llm safety training fail?arXiv preprint arXiv:2307.02483 (2023).
• Wei etal. (2022)Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, QuocV Le, Denny Zhou, etal. 2022.Chain-of-thought prompting elicits reasoning in large language models.Advances in Neural Information Processing Systems 35 (2022), 24824–24837.
• Wu etal. (2023a)Shengqiong Wu, Hao Fei, Leigang Qu, Wei Ji, and Tat-Seng Chua. 2023a.NExT-GPT: Any-to-Any Multimodal LLM.arXiv preprint arXiv:2309.05519 (2023).
• Wu etal. (2023b)Shijie Wu, Ozan Irsoy, Steven Lu, Vadim Dabravolski, Mark Dredze, Sebastian Gehrmann, Prabhanjan Kambadur, David Rosenberg, and Gideon Mann. 2023b.Bloomberggpt: A large language model for finance.arXiv preprint arXiv:2303.17564 (2023).
• Xie etal. (2023)Qianqian Xie, Weiguang Han, Xiao Zhang, Yanzhao Lai, Min Peng, Alejandro Lopez-Lira, and Jimin Huang. 2023.PIXIU: A Large Language Model, Instruction Data and Evaluation Benchmark for Finance.arXiv preprint arXiv:2306.05443 (2023).
• Xu and Cohen (2018)Yumo Xu and ShayB Cohen. 2018.Stock movement prediction from tweets and historical prices. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 1970–1979.
• Yang etal. (2023b)Hongyang Yang, Xiao-Yang Liu, and ChristinaDan Wang. 2023b.FinGPT: Open-Source Financial Large Language Models.arXiv preprint arXiv:2306.06031 (2023).
• Yang etal. (2023a)Jingfeng Yang, Hongye Jin, Ruixiang Tang, Xiaotian Han, Qizhang Feng, Haoming Jiang, Bing Yin, and Xia Hu. 2023a.Harnessing the power of llms in practice: A survey on chatgpt and beyond.arXiv preprint arXiv:2304.13712 (2023).
• Yang etal. (2022)Linyi Yang, Jiazheng Li, Ruihai Dong, Yue Zhang, and Barry Smyth. 2022.NumHTML: Numeric-Oriented Hierarchical Transformer Model for Multi-task Financial Forecasting. In Proceedings of the AAAI Conference on Artificial Intelligence, Vol.36. 11604–11612.
• Yang etal. (2020)Linyi Yang, Tin LokJames Ng, Barry Smyth, and Riuhai Dong. 2020.Html: Hierarchical transformer-based multi-task learning for volatility prediction. In Proceedings of The Web Conference 2020. 441–451.
• Yang etal. (2016)Zichao Yang, Diyi Yang, Chris Dyer, Xiaodong He, Alex Smola, and Eduard Hovy. 2016.Hierarchical attention networks for document classification. In Proceedings of the 2016 conference of the North American chapter of the association for computational linguistics: human language technologies. 1480–1489.
• Yao etal. (2022)Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik Narasimhan, and Yuan Cao. 2022.React: Synergizing reasoning and acting in language models.arXiv preprint arXiv:2210.03629 (2022).
• Yao etal. (2023)Weiran Yao, Shelby Heinecke, JuanCarlos Niebles, Zhiwei Liu, Yihao Feng, Le Xue, Rithesh Murthy, Zeyuan Chen, Jianguo Zhang, Devansh Arpit, etal. 2023.Retroformer: Retrospective large language agents with policy gradient optimization.arXiv preprint arXiv:2308.02151 (2023).
• Ye etal. (2023)Seonghyeon Ye, Hyeonbin Hwang, Sohee Yang, Hyeongu Yun, Yireun Kim, and Minjoon Seo. 2023.In-context instruction learning.arXiv preprint arXiv:2302.14691 (2023).
• Yu etal. (2023)Xinli Yu, Zheng Chen, Yuan Ling, Shujing Dong, Zongyi Liu, and Yanbin Lu. 2023.Temporal Data Meets LLM–Explainable Financial Time Series Forecasting.arXiv preprint arXiv:2306.11025 (2023).
• Zhang etal. (2023)Wenxuan Zhang, Yue Deng, Bing Liu, SinnoJialin Pan, and Lidong Bing. 2023.Sentiment Analysis in the Era of Large Language Models: A Reality Check.arXiv preprint arXiv:2305.15005 (2023).
• Zhang and Gao (2023)Xuan Zhang and Wei Gao. 2023.Towards llm-based fact verification on news claims with a hierarchical step-by-step prompting method.arXiv preprint arXiv:2310.00305 (2023).
• Zhao etal. (2023)WayneXin Zhao, Kun Zhou, Junyi Li, Tianyi Tang, Xiaolei Wang, Yupeng Hou, Yingqian Min, Beichen Zhang, Junjie Zhang, Zican Dong, etal. 2023.A survey of large language models.arXiv preprint arXiv:2303.18223 (2023).
• Zhao etal. (2021)Zihao Zhao, Eric Wallace, Shi Feng, Dan Klein, and Sameer Singh. 2021.Calibrate before use: Improving few-shot performance of language models. In International Conference on Machine Learning. PMLR, 12697–12706.

Appendix A Dataset and Clustering Pipeline

In this section, we include additional details on the statistics of the collected dataset and the overall clustering pipeline.

Appendix B Full Prompt Examples

In this section, we provide full examples of the prompts used in SEP and the responses. Examples for four tasks are shown:

Appendix C Evaluation of Explanation Quality

The explanation of the metrics used in Table 2 are given below. These metrics are manually curated by us through the assistance of ChatGPT. There are currently limited works on evaluating generated stock explanations, given that it is a relatively new application.