Temporal Data Meets LLM -

Explainable Financial Time Series Forecasting

Xinli Yu Zheng Chen Yuan Ling
[email protected] [email protected] [email protected]
Amazon AWS Amazon Alexa AI Amazon Benchmarking
Seattle, WA, USA Seattle, WA, USA Seattle, WA, USA

Shujing Dong Zongyi Liu Yanbin Lu

arXiv:2306.11025v1 [cs.LG] 19 Jun 2023

[email protected] [email protected] [email protected]

Amazon Benchmarking Amazon Benchmarking Amazon Alexa AI
Seattle, WA, USA Seattle, WA, USA Seattle, WA, USA
ABSTRACT KEYWORDS
This paper presents a novel study on harnessing Large Language Time Series, Temporal Data, Large Language Model, Explainable
Models’ (LLMs) outstanding knowledge and reasoning abilities for AI, Multi-Modal Learning
explainable financial time series forecasting. The application of
machine learning models to financial time series comes with sev- ACM Reference Format:
eral challenges, including the difficulty in cross-sequence reason- Xinli Yu, Zheng Chen, Yuan Ling, Shujing Dong, Zongyi Liu, and Yanbin
ing and inference, the hurdle of incorporating multi-modal signals Lu. 2018. Temporal Data Meets LLM - Explainable Financial Time Series
from historical news, financial knowledge graphs, etc., and the is- Forecasting. In Proceedings of Make sure to enter the correct conference title
sue of interpreting and explaining the model results. In this paper, from your rights confirmation emai (Conference acronym ’XX). ACM, New
we focus on NASDAQ-100 stocks, making use of publicly acces- York, NY, USA, 10 pages. https://doi.org/XXXXXXX.XXXXXXX
sible historical stock price data, company metadata, and historical
economic/financial news. We conduct experiments to illustrate the
potential of LLMs in offering a unified solution to the aforemen- 1 INTRODUCTION
tioned challenges. Our experiments include trying zero-shot/few- The rapid advancements in Machine Learning (ML) and Artificial
shot inference with GPT-4 and instruction-based fine-tuning with Intelligence (AI) technologies over the past few years have opened
a public LLM model Open LLaMA. We demonstrate our approach up numerous opportunities and challenges across various domains,
outperforms a few baselines, including the widely applied classic including the realm of financial markets [4, 32, 49]. In particular,
ARMA-GARCH model and a gradient-boosting tree model. Through the task of financial time series forecasting, a key element in strate-
the performance comparison results and a few examples, we find gic decision-making and policy formation, has witnessed signifi-
LLMs can make a well-thought decision by reasoning over infor- cant technological innovations, from statistical/econometric time
mation from both textual news and price time series and extract- series techniques [2, 7, 21, 46], to machine learning techniques
ing insights, leveraging cross-sequence information, and utilizing [31, 33, 68], to deep learning [13, 28, 34, 35, 52]. Despite these ad-
the inherent knowledge embedded within the LLM. Additionally, vancements, there are several inherent challenges associated with
we show that a publicly available LLM such as Open-LLaMA, af- the deployment of ML/AI models in finance.
ter fine-tuning, can comprehend the instruction to generate ex- One challenge lies in the realm of cross-sequence reasoning
plainable forecasts and achieve reasonable performance, albeit rel- and inference, a vital aspect for understanding temporal patterns
atively inferior in comparison to GPT-4. and making accurate predictions. The current approaches include
time-series correlation analysis [8, 11, 20, 48] and clustering [1, 3,
CCS CONCEPTS 50]. Deep learning has recently been leveraged to learn from the
• Computing methodologies → Time series analysis; Natural complex latent dependencies among time series [25, 40, 42, 55].
language processing; Explainable AI . Despite these advancements, existing methods have yet to effec-
tively capture the intricate dependencies characteristic of time se-
ries data. The varying design, implementation, and data require-
ments of these methods further creates a barrier for their wide-
Permission to make digital or hard copies of all or part of this work for personal or
classroom use is granted without fee provided that copies are not made or distributed spread application in the field.
for profit or commercial advantage and that copies bear this notice and the full cita- Another notable hurdle involves handling complex multi-modal
tion on the first page. Copyrights for components of this work owned by others than financial temporal data that extends beyond numeric sequences.
ACM must be honored. Abstracting with credit is permitted. To copy otherwise, or re-
publish, to post on servers or to redistribute to lists, requires prior specific permission The data may encapsulate diverse sources such as historical news,
and/or a fee. Request permissions from [email protected]. financial knowledge graphs, social media activities, and various
Conference acronym ’XX, June 03–05, 2018, Woodstock, NY other market indicators. There has been recent effort leveraging
© 2018 Association for Computing Machinery.
ACM ISBN 978-1-4503-XXXX-X/18/06. . . $15.00 statistical inference [29], RNN/CNN with text embedding [62], graph
https://doi.org/XXXXXXX.XXXXXXX neural networks [9], etc. to integrate the complex information.
Conference acronym ’XX, June 03–05, 2018, Woodstock, NY Xinli Yu, Zheng Chen, Yuan Ling, Shujing Dong, Zongyi Liu, and Yanbin Lu

Lastly, but of utmost importance, the issue of interpretabil- • We compare our proposed LLM approaches with existing
ity and explainability poses significant challenges to the trust- methods, which include an ARMA-GARCH model and a gradient-
worthiness of machine learning and deep learning models. The boosting tree model. We demonstrate even zero-shot infer-
majority of existing deep learning models operate as black boxes, ence using GPT-4 can outperform a boosting-tree model with
offering little insight into their decision-making processes. This about ∼300 features.
lack of transparency sometimes raises concerns about the relia-
bility of their results and impedes user trust. This is particularly 2 RELATED WORKS
relevant in sensitive fields like finance, where substantial invest-
The field of financial time series forecasting has been a subject
ments and assets are at stake. There is recent study trying to un-
of extensive research, with various methodologies being proposed
derstand deep-learning based predictions mainly through the at-
over the years. While traditional statistical methods and machine
tention scores [24], but such insight is still not readily human read-
learning techniques have made significant contributions to this
able and still requires substantial interpretation effort.
field, the advent of LLMs presents new and significant potentials.
The recent advancement of Large Language Models (LLMs) [5,
6, 43, 59] potentially lend us a powerful tool to address all above
challenges in a unified, flexible way. 2.1 Traditional Statistical/Econometric
First, LLMs can learn the complex relations among sequences. Methods
LLMs are so far the most powerful Transformer-based models, and Traditional statistical/econometric methods have long been the cor-
there has been abundant previous researches showing Transformer- nerstone of financial time series forecasting. Techniques such as
based models are capable of learning the underlying complex rela- ARMA-GARCH models have been widely used due to their ability
tions among textual sequences [15, 51, 67, 69, 70] and solving quan- to capture dependencies and volatility clustering in financial time
titative problems [26, 36, 64]. It is reasonable to expect the poten- series [2, 14, 17, 22]. These models have been extended and mod-
tial of LLMs understanding complex dependencies among numeric ified in various ways to better capture the complexities of finan-
time series augmented by temporal textual sequences. cial markets [19, 23, 39, 56]. Other popular statistical/econometric
Secondly, LLMs have demonstrated outstanding reasoning methods for financial time series include Vector Autoregressive
and inference capability over multi-modal data. By design, Models (VAM) [73], State-Space Models and the Kalman Filter [12],
LLMs are proficient at learning from a broad spectrum of data Diffusion Models [16], Vector Error Correction Model (VECM) [27],
sources and types. They are trained on a vast amount of texts from Dynamic Stochastic General Equilibrium (DSGE) [54], etc.
the internet, encompassing a wide range of topics, styles, and for-
mats. This equips them to handle diverse input data, such as numer-
ical, textual, structured data [53, 65]. This multi-modal data han- 2.2 Machine Learning Techniques
dling capability could be particularly useful for financial forecast- With the advent of machine learning, a variety of models have been
ing, where crucial information often comes from disparate sources, applied to financial forecasting. Decision trees, support vector ma-
such as numerical market data, textual news articles, and social me- chines, etc., have been actively studied for financial time series
dia posts. prediction [37, 45, 60, 61, 63, 66]. More recently, deep learning tech-
Lastly, LLMs are natural explainers that generate human niques, such as Recurrent Neural Networks (RNNs), Convolutional
readable explanations providing insight into a decision. One Neural Networks (CNNs), and Transformer models, have been ap-
of the key advantages of LLMs is their ability to generate natu- plied to this task, demonstrating their ability to capture complex,
ral language text that is coherent, contextual, and comprehensive. non-linear relationships in the data [13, 28, 34, 35, 52].
This allows them to provide human-readable explanations for their
decisions [72]. Furthermore, through Chain-of-Thoughts (COT) or 2.3 Large Language Models
step-by-step thinking [38, 64, 71], beyond a few sentences of expla-
The recent development of Large Language Models (LLMs) has
nation, LLMs can even generate detailed step-by-step reasoning to
opened up new possibilities for financial time series forecasting.
reveal the decision-making process.
LLMs, such as GPT-3 [5] and GPT-4 [43], LLaMA[58] (including
The following summarizes the main contributions of this paper,
Alpaca[57], Vincuna[10]), have demonstrated remarkable capabil-
ities in reasoning and understanding complex dependencies in the
heterogeneous data, and the ability to generate human-readable
• This paper takes a novel exploration to study LLMs’ poten-
explanations for their decisions [38, 64, 71, 72]. However, the ap-
tial to the valuable task of explainable financial time series
plication of LLMs in financial time series forecasting with expla-
forecasting. For this paper, we focus on the NASDAQ-100
nation is still a relatively unexplored area, and this paper aims to
stock price time series. To the best of our knowledge, there
contribute to this emerging field.
is not yet public studies on this topic to date.
• We experiment with a combination of zero-shot/few-shot in-
ference techniques with the state-of-the-art AI model GPT-4 3 METHODOLOGY
[43], and instruction-based fine-tuning using Open LLaMA In this study, we focus on the NASDAQ-100 stock price time se-
[18]. Our experiment results also show that the technique ries, supplemented by metadata about the stock company and rel-
of chain-of-thoughts helps boost the performance in most evant financial news data concerning both the specific stock and
of the experiments. the broader financial/economic landscape. Our primary focus is on
Temporal Data Meets LLM -
Explainable Financial Time Series Forecasting Conference acronym ’XX, June 03–05, 2018, Woodstock, NY

forecasting weekly/monthly stock returns (defined as the percent- in Figure 4, includes instructions, the company profile, a histori-
age change in stock price from the beginning to the end of the cal temporal news summary/keywords sequence intermixed with
week/month) with accompanying explanations. This focus aligns the categorized stock price time series, and cross-sequence few-shot
well with the expertise of Large Language Models (LLMs). learning examples.
We demonstrate our structured design of prompts for LLMs and To avoid unnecessary repetition in the prompt text, we pur-
apply the state-of-the-art GPT-4 model [44] for zero-shot and few- posely provide few-shot learning examples from stocks similar to
shot inference tasks. For fine-tuning, we utilize the publicly avail- the subject of interest. This design also assists us in demonstrat-
able Open LLaMA [18]. We also incorporate the technique of Chain ing that the LLM can consider cross-sequence information from
of Thoughts (COT) [38, 64], which has been found to enhance the various stocks. To identify similar stocks, we query GPT-4 with
effectiveness of LLMs in other research studies. a question such as "List the top 3 NASDAQ stocks most similar to
AAPL". A typical response, such as "MSFT, GOOGL, AMZN", show-
3.1 Data cases the LLM’s comprehension of the relationships between vari-
3.1.1 Stock Price Data. We download daily NASDAQ-100 stock 1 ous financial entities and concepts. In employing an LLM, we are
implicitly leveraging its extensive knowledge of financial entities
price data from Yahoo Finance using yfinance package (pypi.org/project/yfinance/).
In this paper, we first normalize the numeric price time series as and concepts.
a percentage-change time series, and then categorize the percent- The prompt structure and instructions have been empirically
age change into bins. For example, for weekly forecasting, we cat- tweaked. For instance, we divided the instruction into two parts,
egorize price change between this week and last week into 12 bins positioning them at the beginning and end of the prompt, which
"D5+", "D5", "D4", "D3", "D2", "D1", "U1", "U2", "U3", "U4", "U5", "U5+", aids the model in better recognizing its task: to predict next week’s
where "D5+" means price dropping more than 5%, "D𝑖" (𝑖=5,4,3,2,1) summary and keywords, rather than summarizing historical data.
means price dropping between (𝑖 − 1)% and 𝑖%, "U5+" means price The predicted summary and keywords serve as the explanation for
rising more than 5%, "U𝑖" (𝑖=1,2,3,4,5) means price rising between the corresponding stock return prediction.
(𝑖 − 1)% and 𝑖%. The number of bins might vary for inference at We also experimented with the Chain-of-Thoughts approach
different granularity. For example, for monthly inference, we al- [38, 64, 71], i.e., the idea of "step-by-step thinking", by appending
low 𝑖 be up to 10, and we have corresponding "D10+" and "U10+" the instruction "Can you reason step by step before finalizing the
categories. output?" to the end of the prompt. To our surprise, this notably
improved the performance by a few points (see Section 4.2). The
3.1.2 Company Profile Data. We use GTP-4 to generate company result of the step-by-step thinking process in response to Figure 4
description, general positive/negative factors that might impact is illustrated in Figure 5, where it is evident that GPT-4 identifies a
the company’s stock price. Figure 1 is an example of the prompt previously overlooked crucial point about "earnings reports" when
to ask GPT-4 to generate the company profile, and the GPT-4 re- explicit reasoning steps are generated.
sponse.
3.1.3 Finance/Economy News Data. We use Google Custom Search 3.3 Instruction-based Fine-tuning with Open
API to obtain stock top-5 news stories on a weekly basis for each LLaMA
of the NASDAQ-100 stocks. After that, we use GPT-4 to generate
We perform instruction-based fine-tuning using Open LLaMA 13B
a summary and extract keywords from each obtained news article.
model to see how well a publicly available model could perform
An example of prompt and GPT-4 response is shown in Figure 2.
in comparison to GPT-4, especially after fine-tuning. The Open
In addition, a similar method is applied to obtain top-5 news
LLaMA 13B model, in its zero-shot inference, typically tends to
stories about macro economy and finance status of each week.
replicate portions of the prompt rather than executing the instruc-
To reduce input size, We further generate meta summary and
tions in the prompt effectively. Therefore, it is incapable of prop-
keywords for each week using GPT-4, given all the top story sum-
erly handling instruction-based prompts as shown in Figure 4 with-
maries and keywords of the week. We only use the meta summary
out undergoing a process of fine-tuning. Therefore we focus on
and the keywords in this paper’s experiments. An example of meta
fine-tuning with the Open LLaMA model in this paper.
summary and keywords is shown in Figure 3; they look similar to
Instruction-based fine-tuning has been recently shown to be ef-
the example in Figure 2, but much condensed.
fective in guiding the model’s training process with specific direc-
tives [47, 57]. We created a dataset of 30K weekly forecasting plus
3.2 Instruction-Based Zero-shot/Few-shot 7K monthly forecasting, derived from 5-year historical data span-
Inference with LLMs ning from Jun 2017 to June 2022. Unlike GPT-4 that supports up to
In zero-shot and few-shot inference, LLMs demonstrate their abil- 8K token size, we need to compress the prompt into 1K tokens for
ity to generate responses either without any additional examples fine-tuning Open LLaMA, due to model and hardware constraints.
(zero-shot) or based on a minimal number of examples beyond the For each fine-tuning example, we employ GPT-4 to condense the
original training set (few-shot). full historical meta news summary/keywords (e.g. from week 8 to
In our zero-shot/few-shot inference experiment, we utilize an the last week as shown in Figure 4) into a single, even more con-
instruction-based prompt. The structure of our prompt, illustrated cise summary/keywords pair. Simultaneously, the "Company Pro-
file" and "Forecasting Examples" sections of the prompt are also
1 Using the stock list of Jun 01, 2023 respectively condensed into more succinct summary paragraphs.
Conference acronym ’XX, June 03–05, 2018, Woodstock, NY Xinli Yu, Zheng Chen, Yuan Ling, Shujing Dong, Zongyi Liu, and Yanbin Lu

Prompt: Generate a short description for stock AAPL’ s company. Also list general positive and negative factors that might
impact the stock price; be brief and use keywords. Consider diverse general factors, such as macro economic situation (e.g.
inflation, CPI growth), business factors (e.g. sales, investment, products), technology factors (e.g. innovation), and others. Use
format Description: ..., Positive Factors: ..., Negative factors: ...

GPT-4 Response:
Description:
AAPL (Apple Inc.) is a world-renowned technology company based in Cupertino, California. Known for its sleek designs and innovative
technologies, the company produces a wide range of products including the iPhone, iPad, Mac, Apple Watch, and services like iCloud, App Store,
and more.

Positive Factors:
Macroeconomic: Economic growth, increased consumer spending
Business: Strong sales, expansion into new markets, positive earnings reports, strategic acquisitions
Technology: Innovation, new product launches, advancements in artificial intelligence and augmented reality
Brand: Strong global brand, loyal customer base
Legal/Regulatory: Favorable policy changes, successful patent applications

Negative Factors:
Macroeconomic: Inflation, economic downturn, trade restrictions
Business: Decreased sales, poor earnings reports, lack of new investments
Technology: Slow pace of innovation, delays in product launches
Brand: Damage to brand reputation, loss of customer loyalty
Legal/Regulatory: Antitrust investigations, unfavorable policy changes, patent disputes
Environmental: Failure to meet sustainability goals, negative environmental impact

Figure 1: An example of a stock’s company profile consisting of the company description, the general positive/negative factors
impacting the company’s stock prices.

Prompt [template, those inside {} are placeholders]: Please summarize the following noisy but possible news data extracted from
web page HTML, and extract keywords of the news. The news text can be very noisy due to it is HTML extraction. Give formatted
answer such as Summary: ..., Keywords: ... The news is supposed to be for {symbol} stock. You may put ’N/A’ if the noisy text does
not have relevant information to extract.

News: {news}

GPT-4 Response:
Summary: Apple Inc is postponing perks for some corporate divisions and expanding a cost-cutting effort, joining its peers in streamlining oper-
ations. The company proposes to forego bonuses or promotions for April, with the payments occurring only in October. Apple is also restricting
hiring and leaving additional positions open after employee departure. This is triggering fear of downsizing among employees.
Keywords: Apple, bonus, hiring, cost-cutting, promotions, downsizing, employees

Figure 2: An example of news summary and keywords extracted from one news article for a stock’s company (AAPL in this
case), generated by GPT-4. The original news is at https://sports.yahoo.com/apple-joins-cost-cut-bandwagon-145845685.html.
The prompt in this example is a template. Texts inside [] are comments and are not really in the prompt when we submit it to
the LLM, and we will fill stock symbol and the news into the placeholders enclosed by {}.

While it would be ideal for Open LLaMA to manage its own end- and keywords extracted by Open LLaMA 13B model often fall short
to-end experiment, including the task of prompt compression for of usability.
fine-tuning, we still resort to using GPT-4 right now. This is due
to Open LLaMA 13B model’s zero-shot summarization capability is
considerably inferior to those of GPT-4 in practice. The summaries
Temporal Data Meets LLM -
Explainable Financial Time Series Forecasting Conference acronym ’XX, June 03–05, 2018, Woodstock, NY

Summary: Apple Inc.’s stock is displaying rising relative strength, although shy of a key benchmark, and the company has been
highlighted as a top-performing stock due to its sales of 1.3 billion iPhones and a significant $19.4 billion profit last quarter. As
the NASDAQ surges with Apple as a big cap, Apple expands cost-cutting measures, postpones perks, restricts hiring, and leaves
positions open, instigating downsizing concerns. The overall stock market performance is boosted by a $30 billion deposit infusion
for struggling firms.
Keywords: Apple Inc., stock, relative strength, cost-cutting, downsizing, NASDAQ, $30 billion deposit infusion, iPhones, profit.

Figure 3: An example of one week’s meta summary and keywords condensed from all the company’s summaries and keywords
from the week.

Once fine-tuned, the Open LLaMA 13B model demonstrates a 4.1.3 Evaluation Metrics. We perform weekly and monthly stock
much more satisfactory comprehension of the instruction, result- return forecasting with the baselines and LLM-based methods. We
ing in the generation of a forecast and an accompanying explana- treat 4 weeks as one month for convenience, and therefore there
tion that appears coherent. This is illustrated in Figure 6. As per are 13 "month"s in the 52-week evaluation time window.
the result in section 4.2, when it comes to binary classification, To evaluate the performance of our forecasting models, we em-
the Open LLaMA model’s performance is competitive compared ploy three metrics.
to GPT-4. However, we’ve noticed that the Open LLaMA model • Binary precision assesses the model’s ability to correctly
has a tendency to produce more extreme predictions, such as U5+ predict the general direction of stock price movement, i.e.,
or D5+, which result in a relatively higher squared error. "Up" (U) or "Down" (D).
• Bin precision, on the other hand, evaluates the model’s ac-
4 EXPERIMENTS curacy in predicting the exact bin from a full list of bins such
4.1 Experiment Setup as "D5+", "D5", "D4", ..., "D1", "U1", ..., "U5", "U5+".
• The MSE of consecutive bin ordinals (e.g., -6 for "D5+", -5 for
4.1.1 Data Time Window. The details of the data used in the ex- "D5", ..., 0 for "U1", ..., 4 for "U5", 5 for "U5+") is used to mea-
periments is as described in Section 3.1. We focus on NASDAQ-100 sure the average squared differences between the model’s
stock return forecasting for this paper.
predictions and the actual values. This metric helps to un-
• The training/fine-tuning time window contains 5-year data derstand the model’s tendency to make extreme forecasts
from Jun 12 2017 to Jun 05 2022. The data in this time win- when its predictions are incorrect.
dow is used for training of the baseline models, and the fine- To evaluate the quality of the forecasting explanation (the pre-
tuning of the Open LLaMA 13B model. dicted next-week/month summary/keywords), we employ ROGUE-
• The evaluation time window has 52 weeks spanning from 1 and ROGUE-2 scores to compare with the actual summary/keywords
Jun 06 2022 to Jun 04 2023. The evaluation of baseline mod- by GPT-4 extracted from the actual top news of the next week/month.
els, the zero/few-shot inference experiments with GPT-4,
and the evaluation of fine-tuned Open LLaMA 13B model,
4.2 Performance Evaluation
are based on data in this time window.
The results of our experiments are summarized in Table 1 and 2. Ta-
4.1.2 Baseline Models. To evaluate the performance of our approach, ble 1 provides a comparative analysis of our LLM-based methods
we include a heuristic baseline using the most-frequent historical and the baseline models in terms of their performance in forecast-
bin (i.e. the most frequent bin from historical weeks before the tar- ing stock returns. Table 2, on the other hand, evaluates the quality
get week to forecast) as the prediction, an ARMA-GARCH model of the explanations generated by the LLMs.
(𝑝 = 𝑞 = 1) [39, 56], and a gradient-boosting tree model [41] im- In summary, our results show the effectiveness of LLMs in fi-
plemented by LightGBM package [30]. These baseline models are nancial time series forecasting, with "GPT-4 few-shot with COT"
trained on the training/fine-tuning data time window, and evalu- consistently showing the best performance in both prediction ac-
ated on the evaluation time window. curacy and explanation quality. The results also highlight the tech-
For the gradient-boosting tree model, we include the following nique of Chain-of-Thoughts (COT) consistently boosts performance,
features. There are total about 300 features for the tree. and the potential of instruction-based fine-tuning with publicly
available LLMs like Open LLaMA to achieve reasonable perfor-
(1) Historical price time series available in the daily stock price mance in comparison to GPT-4 through fine-tuning with COT.
data, including open, close, min, max prices, and the daily
trading volume. 4.2.1 Stock Price Forecasting. From the results of Table 1, we ob-
(2) The average, medium, min, max, and stddev of a rolling win- serve that both GPT-4 and Open LLaMA 13B model outperform
dow of size 2, 5, 10, 30, 60, 90 for the above time series. the ARMA-GARCH model and the gradient-boosting tree model in
(3) The stock sector information, and the stock historical earn- terms of both binary and bin precision. GPT-4, in particular, shows
superior performance in both zero-shot and few-shot settings, with
ing obtained from Alpha Vantage (https://www.alphavantage.co/documentation/).
Conference acronym ’XX, June 03–05, 2018, Woodstock, NY Xinli Yu, Zheng Chen, Yuan Ling, Shujing Dong, Zongyi Liu, and Yanbin Lu

Prompt [template, those inside {} are placeholders]:

Instruction: Forecast next week stock return (price change) for symbol, given the company profile, historical weekly news summary,
keywords, and stock returns, and optionally the examples from other stocks of a similar company.

The trend is represented by bins "D5+", "D5", "D4", "D3", "D2", "D1", "U1", "U2", "U3", "U4", "U5", "U5+", where "D5+" means price
dropping more than 5%, D5 means price dropping between 4% and 5%, "D4" means price dropping between 3% and 4%, "U5+" means
price rising more than 5%, "U5" means price rising between 4% and 5%, "D4" means price rising between 3% and 4%, etc.
Company Profile: {company_profile} [company description, general positive/negative factors as shown in Figure 1.]
Recent News: News are ordered from oldest news to latest news.
8 Weeks ago.
{meta_news_summary_company} [meta news summary and keywords of the week for the company]
{meta_news_summary_macro} [meta news summary and keywords for the macro economy and finance status of the week]
...
7 weeks ago.
...
Last week.
...
Forecasting Examples: {few_shot_learning_examples_from_similar_stocks} [We use forecasting examples from top-3 similar stocks
as the few-short learning examples. The example is structured similar to this prompt, with company profile and recent news, but
without the instruction texts.]

Now predict what could be the next week’s Summary, Keywords, and forecast the Stock Return. The predicted Summary/Keywords
should explain the stock return forecasting. You should predict what could happen next week. Do not just summarize the his-
tory. The next week stock return need not be the same as the previous week. Use format Summary: ..., Keywords: ..., Stock Return: ...

GPT-4 Response [Forecasting week 05/01/2023 to 05/07/2023]:

Summary: Apple’s stock is expected to see fluctuations in the week to come against a backdrop of mixed news. The company’s announced
expansion into India, coupled with the introduction of high-yield savings accounts, initially provided a boost to the stock. However, growing
concerns surrounding reduced PC shipments and apprehensions about the company’s upcoming earnings injected a degree of uncertainty. Despite
these challenges, Apple’s unwavering focus on cost control and the anticipation surrounding innovative offerings, such as the forthcoming
mixed-reality headset, shored up investor confidence. Consequently, the stock maintained a general upward trend throughout the week, albeit
punctuated by a few dips. The broader economic climate on Wall Street also had a role to play, with a recovering sentiment in Q1 towards
high-tech companies like Microsoft and Alphabet. These companies released pivotal financial reports and ground-breaking products during the
period, influencing the tech sector’s performance, potentially impacting Apple’s stock trajectory.

Keywords: AAPL, mixed-reality headset, India expansion, high-yield savings account, cost control strategies, Bernstein, Wall Street, economic
recovery, tech sector performance

Stock Return: U1

Ground Truth [Summary from actual news from 05/01/2023 to 05/07/2023 and the actual return]:
Summary: Apple’s stock saw significant gains after outperforming Wall Street’s fiscal second quarter targets, driven by robust iPhone sales and
strong services, despite posting a decrease in revenue for the second consecutive quarter and facing potential competition in the VR headset
market from Reliance Jio’s JioDive. The company also noted an improvement in supply and strength in emerging markets, further boosting
investor confidence.

Keywords: Apple, stock, Wall Street, fiscal second quarter, iPhone, services sales, quarterly earnings, revenue, declining revenue, emerging markets,
improved supply, Virtual Reality, competition, JioDive, Reliance Jio.

Stock Return: U3

Figure 4: The prompt structure for the experiments in this paper with LLMs. We also give an example of GPT-4 response to
a concrete prompt constructed from information on and before 04/30/2023. We notice the cross-sequence information and
macro-economy information are obviously considered in the LLM’s reasoning. The stock return forecast U1 is for the next
week from 05/01/2023 to 05/07/2023. The actual market performance for AAPL is U3. We note Apple 2023 Q2 earning call
happened on May 04 beat expectation, which maybe the major contributor to the higher-than-forecasting gain in the week.
Temporal Data Meets LLM -
Explainable Financial Time Series Forecasting Conference acronym ’XX, June 03–05, 2018, Woodstock, NY

Prompt: {The same prompt as in Figure 4.}

Can you reason step by step before the finalized output?

GPT-4 Response [Forecasting week 05/01/2023 to 05/07/2023]:

Given the information provided, we can see that Apple (AAPL) has had a number of positive developments and a few potential concerns in the
recent past.
Positive Developments:
• Recent sales and earnings have been strong, and the company appears to be exploring new markets such as India.
• The upcoming mixed-reality headset and the "Apple Pay Later" service show that Apple is still innovating and pushing into new product
areas.
• The cost-cutting measures and the plan to use 100% recycled cobalt show that Apple is taking steps to improve its efficiency and sustainability,
which could be attractive to certain investors.
• Winning the lawsuit against Britain’s antitrust regulator and avoiding potential penalties.
• AAPL stock has been trending upwards recently, with multiple weeks of upward movement (U5, U4, U3, U2), and only a single day of slight
downward movement (D1).
Potential Concerns:
• Apple is cutting costs, delaying bonuses, and slowing down hiring in certain areas, which could potentially affect employee morale and
productivity.
• The negative sentiment from analyst Toni Sacconaghi about the overestimation of Q3 and Q4 revenues by Wall Street could potentially
impact the stock price.

In addition to above factors, the general direction of the stock market appears to be on an upward trajectory, as evidenced by the recent increases
in the Dow Jones, S&P 500, and Nasdaq. In recent weeks, the stock performances of high-tech giants such as Microsoft and Alphabet have also
shown considerable strength. Furthermore, recent positive developments in the macroeconomic environment, including a surge in consumer
spending, seem encouraging.

Now, we can generate a prediction for next week’s Summary, Keywords, and Stock Return. However, it’s important to note that while the
information provided is used as a basis, this is a prediction and the actual outcome may be different.

Summary: Apple’s mixed-reality headset previews have received positive feedback from industry experts, increasing investor confidence. However,
apprehension about cost-cutting measures and fears of slower hiring might add some uncertainty. Wall Street anticipates a strong earnings report,
boosting stock morale. Furthermore, Apple’s expansion into India continues to show promise.

Keywords: Apple, mixed-reality headset, investor confidence, cost-cutting, hiring, Wall Street, earnings report, expansion, India.

Stock Return: U2. Given the positive trends and promising developments, the stock is likely to experience upward movement. However, cost-cutting
concerns may moderate gains, leading to a prediction of "U2", suggesting a rise of 1% 2% next week.

Figure 5: GPT-4 outputs its detailed reasoning steps if we simply add an instruction "Can you reason step by step before the
finalized output?" to the end of the prompt in Figure 4. With detailed reasoning steps, GPT-4 captures a previously missed
point "Wall Street anticipates a strong earnings report, boosting stock morale", and amends the stock return forecast as U2.

Summary: The stock market is likely to react favorably to Apple’s continued focus on cost optimization and revenue generation
through new product launches and market expansions. Despite facing regulatory challenges, the company’s efforts towards envi-
ronmental responsibility and diversification of financial offerings are likely to have a positive impact on its long-term prospects.
As a result, we expect Apple’s stock to see another strong week of growth, with a projected return of U5+.
Keywords: Cost-Cutting, New Product Launches, Expansion, Regulation, Supply Chain
Stock Return: U5+

Figure 6: Explainable forecasting by fine-tuned Open LLaMA 13B model.

Conference acronym ’XX, June 03–05, 2018, Woodstock, NY Xinli Yu, Zheng Chen, Yuan Ling, Shujing Dong, Zongyi Liu, and Yanbin Lu

the few-shot setting with COT achieving the best performance. In series, stock trading volumes, and social network data, and 3) ex-
terms of MSE, "GPT-4 few-shot with COT" also achieves the low- ploring the fine-tuning of larger publicly available models, such as
est error, indicating that it not only best predicts the direction of a 30B model, to enhance reasoning capabilities.
the price change but also provides a more accurate estimate of the
magnitude of the change. REFERENCES
Open LLaMA 13B model, after fine-tuning, shows competitive [1] Saeed Aghabozorgi, Ali Seyed Shirkhorshidi, and Teh Ying Wah. 2015. Time-
performance compared to GPT-4 in terms of binary precision. How- series clustering–a decade review. Information systems 53 (2015), 16–38.
[2] Torben Gustav Andersen, Richard A Davis, Jens-Peter Kreiß, and Thomas V
ever, its bin precision is obviously worse, indicating it lacks com- Mikosch. 2009. Handbook of financial time series. Springer Science & Business
petitive fine-grained reasoning capability to pick the right bin. It Media.
also tends to produce more extreme predictions, resulting in a rel- [3] M Suresh Babu, N Geethanjali, and B Satyanarayana. 2012. Clustering approach
to stock market prediction. International Journal of Advanced Networking and
atively higher MSE. Applications 3, 4 (2012), 1281.
[4] Arash Bahrammirzaee. 2010. A comparative survey of artificial intelligence ap-
plications in finance: artificial neural networks, expert system and hybrid intel-
4.2.2 Explanation Quality. Table 2 shows the quality of the ex- ligent systems. Neural Computing and Applications 19, 8 (2010), 1165–1195.
planations generated by the LLMs (GPT-4 and fine-tuned Open [5] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan,
Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda
LLaMA), evaluated using ROUGE-1 and ROUGE-2 scores for both Askell, et al. 2020. Language models are few-shot learners. Advances in neural
the summary (S) and keywords (K) of the news. information processing systems 33 (2020), 1877–1901.
Again, the results show that "GPT-4 few-shot with COT" achieves [6] Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan,
Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda
the highest ROUGE scores, indicating that it generates the most rel- Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, T. J. Henighan,
evant and accurate explanations for the predictions. Open LLaMA, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeff Wu, Clemens Winter,
after fine-tuning with COT, also shows reasonable explanation qual- Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Ben-
jamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford,
ity in parallel with GPT-4 results without COT. Ilya Sutskever, and Dario Amodei. 2020. Language Models are Few-Shot Learn-
ers. ArXiv abs/2005.14165 (2020).
[7] Cathy WS Chen, Feng-Chi Liu, and Mike KP So. 2011. A review of threshold
5 CONCLUSION time series models in finance. Statistics and its Interface 4, 2 (2011), 167–181.
[8] Zheng Chen, Xinli Yu, Yuan Ling, Bo Song, Wei Quan, Xiaohua Hu, and Erjia
In this paper, we explored the application of Large Language Mod- Yan. 2018. Correlated anomaly detection from large streaming data. In 2018
els (LLMs) in the field of financial time series forecasting, with IEEE International Conference on Big Data (Big Data). IEEE, 982–992.
[9] Dawei Cheng, Fangzhou Yang, Sheng Xiang, and Jin Liu. 2022. Financial time se-
a particular focus on NASDAQ-100 stock return prediction. We ries forecasting with multi-modality graph neural network. Pattern Recognition
demonstrated how LLMs, specifically GPT-4 and Open LLaMA, can 121 (2022), 108218.
[10] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhanghao Wu, Hao Zhang,
be utilized to generate forecasts and provide human-readable ex- Lianmin Zheng, Siyuan Zhuang, Yonghao Zhuang, Joseph E. Gonzalez, Ion Sto-
planations for their predictions. Our approach involved the use of ica, and Eric P. Xing. 2023. Vicuna: An Open-Source Chatbot Impressing GPT-4
structured prompts, which included company profile data, histori- with 90%* ChatGPT Quality. https://vicuna.lmsys.org
[11] Thomas Conlon, Heather J Ruskin, and Martin Crane. 2009. Cross-correlation
cal stock price data, and financial news data, to guide the LLMs in dynamics in financial time series. Physica A: Statistical Mechanics and its Appli-
their forecasting tasks. cations 388, 5 (2009), 705–714.
Our experimental results revealed that LLMs could surpass tra- [12] Piet De Jong and Ben Zehnwirth. 1983. Claims reserving, state-space models
and the Kalman filter. Journal of the Institute of Actuaries 110, 1 (1983), 157–181.
ditional statistical models and machine learning techniques, such [13] Alexiei Dingli and Karl Sant Fournier. 2017. Financial time series forecasting-a
as the ARMA-GARCH model and a gradient-boosting tree model, deep learning approach. International Journal of Machine Learning and Comput-
ing 7, 5 (2017), 118–122.
in performance. Notably, the integration of a step-by-step reason- [14] Feike C Drost and Theo E Nijman. 1993. Temporal aggregation of GARCH pro-
ing process, inspired by the Chain of Thought (COT) approach, cesses. Econometrica: Journal of the Econometric Society (1993), 909–927.
significantly enhanced the performance of our LLM-based models. [15] Vijay Prakash Dwivedi and Xavier Bresson. 2020. A generalization of trans-
former networks to graphs. arXiv preprint arXiv:2012.09699 (2020).
Moreover, our fine-tuning experiments with Open LLaMA demon- [16] Jianqing Fan. 2005. A selective overview of nonparametric methods in financial
strated the feasibility of effectively tuning publicly available LLMs econometrics. Statist. Sci. (2005), 317–337.
for this task, thereby addressing the inherent challenges of cross- [17] Christian Francq and Jean-Michel Zakoian. 2004. Maximum likelihood estima-
tion of pure GARCH and ARMA-GARCH processes. Bernoulli 10, 4 (2004), 605–
sequence reasoning, multi-modal signals integration, and result in- 637.
terpretability in the financial sector. [18] Xinyang Geng and Hao Liu. 2023. OpenLLaMA: An Open Reproduction of LLaMA.
https://github.com/openlm-research/open_llama
In conclusion, our preliminary exploration into the application [19] M Ghahramani and A Thavaneswaran. 2006. Financial applications of ARMA
of LLMs in financial forecasting yielded promising initial results. models with GARCH errors. The Journal of Risk Finance 7, 5 (2006), 525–543.
Despite being in the early stages, the encouraging outcomes pro- [20] Parameswaran Gopikrishnan, Vasiliki Plerou, Yan Liu, LA Nunes Amaral, Xavier
Gabaix, and H Eugene Stanley. 2000. Scaling and correlation in financial time
vide a strong motivation for further exploration in this direction. series. Physica A: Statistical Mechanics and its Applications 287, 3-4 (2000), 362–
As we continue to expand the application of LLMs in the financial 373.
domain, we envision a future where financial forecasting is not [21] Wolfgang Härdle, Helmut Lütkepohl, and Rong Chen. 1997. A review of nonpara-
metric time series analysis. International statistical review 65, 1 (1997), 49–72.
only more precise but also more comprehensible and transparent. [22] Jan S Henneke, Svetlozar T Rachev, Frank J Fabozzi, and Metodi Nikolov. 2011.
This development could significantly contribute to the transforma- MCMC-based estimation of Markov Switching ARMA–GARCH models. Applied
Economics 43, 3 (2011), 259–271.
tion of financial decision-making across the entire sector. [23] Altaf Hossain and Mohammed Nasser. 2011. Comparison of the finite mixture
Future research will dive deeper into these methodologies, in- of ARMA-GARCH, back propagation neural networks and support-vector ma-
cluding but not limited to: 1) extending the studies to include more chines in forecasting financial returns. Journal of Applied Statistics 38, 3 (2011),
533–551.
stock indexes such as SP500 and Russell 2000, 2) integrating the [24] Tsung-Yu Hsieh, Suhang Wang, Yiwei Sun, and Vasant Honavar. 2021. Explain-
research with more data types, such as the macro economy time able multivariate time series classification: a deep neural network which learns
Temporal Data Meets LLM -
Explainable Financial Time Series Forecasting Conference acronym ’XX, June 03–05, 2018, Woodstock, NY

Weekly Monthly (Every 4 Weeks)

Binary Precision Bin Precision MSE Binary Precision Bin Precision MSE
Most-Frequent Historical Bin 50.7% 16.4% 43.5 51.4% 17.2% 155.1
ARMA-GARCH 52.4% 11.1% 22.1 50.5% 6.2% 90.1
Gradient Boosting Tree Model 60.8% 26.4% 24.3 56.4% 17.7% 85.6
GPT-4 Zero-Shot 64.5% 31.2% 20.5 64.8% 26.0% 60.1
GPT-4 Few-Shot 65.8% 32.7% 20.6 65.3% 26.5% 58.2
GPT-4 Few-Shot w/ COT 66.5% 35.2% 18.7 69.5% 28.6% 50.4
Open LLaMA (13B) Fine-Tuned 62.2% 26.5% 23.3 60.1% 22.6% 63.3
Open LLaMA (13B) Fine-Tuned w/ COT 64.7% 30.7% 21.0 62.2% 24.4% 63.5
Table 1: Performance comparison between the baseline models and LLMs for stock price weekly/monthly forecasting.

Weekly Monthly (Every 4 Weeks)

ROUGE-1 (S) ROUGE-2 (S) ROUGE-1 (K) ROUGE-2 (K) ROUGE-1 (S) ROUGE-2 (S) ROUGE-1 (K) ROUGE-2 (K)
GPT-4 Zero-Shot 0.2212 0.0675 0.1295 0.0447 0.2528 0.0665 0.1335 0.0657
GPT-4 Few-Shot 0.2242 0.0526 0.1304 0.0454 0.2450 0.0634 0.1348 0.0644
GPT-4 Few-Shot w/ COT 0.2414 0.0543 0.2083 0.0869 0.2645 0.0758 0.2450 0.1025
Open LLaMA (13B) Fine-Tuned 0.2053 0.0395 0.0927 0.0324 0.2242 0.0474 0.1167 0.0520
Open LLaMA (13B) Fine-Tuned w/ COT 0.2371 0.0434 0.1123 0.0425 0.2436 0.0536 0.1356 0.0834
Table 2: Explanation quality evaluation using ROGUE scores, using the GPT-4 summary/keyword extraction of each week’s
true top news from google search as the ground truth.

to attend to important variables as well as time intervals. In Proceedings of the [40] Romit Maulik, Arvind Mohan, Bethany Lusch, Sandeep Madireddy, Prasanna
14th ACM international conference on web search and data mining. 607–615. Balaprakash, and Daniel Livescu. 2020. Time-series learning of latent-space dy-
[25] Yuxiu Hua, Zhifeng Zhao, Rongpeng Li, Xianfu Chen, Zhiming Liu, and Hong- namics for reduced-order model closure. Physica D: Nonlinear Phenomena 405
gang Zhang. 2019. Deep learning with long short-term memory for time series (2020), 132368.
prediction. IEEE Communications Magazine 57, 6 (2019), 114–119. [41] Alexey Natekin and Alois Knoll. 2013. Gradient boosting machines, a tutorial.
[26] Shima Imani, Liang Du, and Harsh Shrivastava. 2023. Mathprompter: Mathe- Frontiers in neurorobotics 7 (2013), 21.
matical reasoning using large language models. arXiv preprint arXiv:2303.05398 [42] Nam Nguyen and Brian Quanz. 2021. Temporal latent auto-encoder: A method
(2023). for probabilistic multivariate time series forecasting. In Proceedings of the AAAI
[27] Søren Johansen. 1995. Likelihood-based inference in cointegrated vector autore- Conference on Artificial Intelligence, Vol. 35. 9117–9125.
gressive models. OUP Oxford. [43] OpenAI. 2023. GPT-4 Technical Report. ArXiv abs/2303.08774 (2023).
[28] Norberto Ritzmann Júnior and Julio Cesar Nievola. 2018. A generalized finan- [44] OpenAI. 2023. GPT-4 Technical Report. arXiv:2303.08774 [cs.CL]
cial time series forecasting model based on automatic feature engineering using [45] Ping-Feng Pai and Chih-Sheng Lin. 2005. A hybrid ARIMA and support vector
genetic algorithms and support vector machine. In 2018 International Joint Con- machines model in stock price forecasting. Omega 33, 6 (2005), 497–505.
ference on Neural Networks (IJCNN). IEEE, 1–8. [46] Andrew J Patton. 2012. A review of copula models for economic time series.
[29] Nont Kanungsukkasem and Teerapong Leelanupab. 2019. Financial latent Journal of Multivariate Analysis 110 (2012), 4–18.
Dirichlet allocation (FinLDA): Feature extraction in text and data mining for [47] Baolin Peng, Chunyuan Li, Pengcheng He, Michel Galley, and Jianfeng Gao. 2023.
financial time series prediction. IEEE Access 7 (2019), 71645–71664. Instruction tuning with gpt-4. arXiv preprint arXiv:2304.03277 (2023).
[30] Guolin Ke, Qi Meng, Thomas Finley, Taifeng Wang, Wei Chen, Weidong Ma, [48] Vasiliki Plerou, Parameswaran Gopikrishnan, Bernd Rosenow, Luís A Nunes
Qiwei Ye, and Tie-Yan Liu. 2017. Lightgbm: A highly efficient gradient boosting Amaral, and H Eugene Stanley. 1999. Universal and nonuniversal properties
decision tree. Advances in neural information processing systems 30 (2017). of cross correlations in financial time series. Physical review letters 83, 7 (1999),
[31] Kyoung-jae Kim. 2003. Financial time series forecasting using support vector 1471.
machines. Neurocomputing 55, 1-2 (2003), 307–319. [49] Yuan Qi and Jing Xiao. 2018. Fintech: AI powers financial services to improve
[32] Boris Kovalerchuk and Evgenii Vityaev. 2005. Data mining for financial applica- people’s lives. Commun. ACM 61, 11 (2018), 65–69.
tions. Data Mining and Knowledge Discovery Handbook (2005), 1203–1224. [50] Sangeeta Rani and Geeta Sikka. 2012. Recent techniques of clustering of time se-
[33] Bjoern Krollner, Bruce J Vanstone, Gavin R Finnie, et al. 2010. Financial time ries data: a survey. International Journal of Computer Applications 52, 15 (2012).
series forecasting with machine learning techniques: a survey.. In ESANN. [51] Yu Rong, Yatao Bian, Tingyang Xu, Weiyang Xie, Ying Wei, Wenbing Huang,
[34] Pedro Lara-Benítez, Manuel Carranza-García, and José C Riquelme. 2021. An and Junzhou Huang. 2020. Self-supervised graph transformer on large-scale
experimental review on deep learning architectures for time series forecasting. molecular data. Advances in Neural Information Processing Systems 33 (2020),
International Journal of Neural Systems 31, 03 (2021), 2130001. 12559–12571.
[35] Tim Leung and Theodore Zhao. 2021. Financial time series analysis and fore- [52] Omer Berat Sezer, Mehmet Ugur Gudelek, and Ahmet Murat Ozbayoglu. 2020.
casting with Hilbert–Huang transform feature generation and machine learning. Financial time series forecasting with deep learning: A systematic literature re-
Applied Stochastic Models in Business and Industry 37, 6 (2021), 993–1016. view: 2005–2019. Applied soft computing 90 (2020), 106181.
[36] Aitor Lewkowycz, Anders Andreassen, David Dohan, Ethan Dyer, Henryk [53] Li Shen, Yan Sun, Zhiyuan Yu, Liang Ding, Xinmei Tian, and Dacheng Tao. 2023.
Michalewski, Vinay Ramasesh, Ambrose Slone, Cem Anil, Imanol Schlag, Theo On Efficient Training of Large-Scale Deep Learning Models: A Literature Review.
Gutman-Solo, et al. 2022. Solving quantitative reasoning problems with lan- arXiv preprint arXiv:2304.03589 (2023).
guage models. arXiv preprint arXiv:2206.14858 (2022). [54] Frank Smets and Raf Wouters. 2003. An estimated dynamic stochastic general
[37] Wei Li and Jian Liao. 2017. A comparative study on trend forecasting approach equilibrium model of the euro area. Journal of the European economic association
for stock price time series. In 2017 11th IEEE International Conference on Anti- 1, 5 (2003), 1123–1175.
counterfeiting, Security, and Identification (ASID). IEEE, 74–78. [55] Wen Song and Shigeru Fujimura. 2021. Capturing combination patterns of long-
[38] Hunter Lightman, Vineet Kosaraju, Yura Burda, Harri Edwards, Bowen Baker, and short-term dependencies in multivariate time series forecasting. Neurocom-
Teddy Lee, Jan Leike, John Schulman, Ilya Sutskever, and Karl Cobbe. 2023. Let’s puting 464 (2021), 72–82.
Verify Step by Step. arXiv preprint arXiv:2305.20050 (2023). [56] Him Tang, Kai-Chun Chiu, and Lei Xu. 2003. Finite mixture of ARMA-GARCH
[39] Junmei Ma and Xinli Yu. 2013. Research on the Pricing of the Basket Credit model for stock price prediction. In Proceedings of the Third International Work-
Default Swap. Mathematical Computation 2, 4 (2013). shop on Computational Intelligence in Economics and Finance (CIEF’2003), North
Carolina, USA. 1112–1119.
Conference acronym ’XX, June 03–05, 2018, Woodstock, NY Xinli Yu, Zheng Chen, Yuan Ling, Shujing Dong, Zongyi Liu, and Yanbin Lu

[57] Rohan Taori, Ishaan Gulrajani, Tianyi Zhang, Yann Dubois, Xuechen 2023. Bloomberggpt: A large language model for finance. arXiv preprint
Li, Carlos Guestrin, Percy Liang, and Tatsunori B. Hashimoto. arXiv:2303.17564 (2023).
2023. Stanford Alpaca: An Instruction-following LLaMA model. [66] Haiqin Yang, Laiwan Chan, and Irwin King. 2002. Support vector machine
https://github.com/tatsu-lab/stanford_alpaca. regression for volatile stock market prediction. In Intelligent Data Engineering
[58] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne and Automated Learning—IDEAL 2002: Third International Conference Manch-
Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal ester, UK, August 12–14, 2002 Proceedings 3. Springer, 391–396.
Azhar, et al. 2023. Llama: Open and efficient foundation language models. arXiv [67] Chengxuan Ying, Tianle Cai, Shengjie Luo, Shuxin Zheng, Guolin Ke, Di He,
preprint arXiv:2302.13971 (2023). Yanming Shen, and Tie-Yan Liu. 2021. Do transformers really perform badly
[59] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne for graph representation? Advances in Neural Information Processing Systems 34
Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal (2021), 28877–28888.
Azhar, Aur’elien Rodriguez, Armand Joulin, Edouard Grave, and Guillaume Lam- [68] Paul D Yoo, Maria H Kim, and Tony Jan. 2005. Machine learning techniques and
ple. 2023. LLaMA: Open and Efficient Foundation Language Models. ArXiv use of event information for stock market prediction: A survey and evaluation. In
abs/2302.13971 (2023). International Conference on Computational Intelligence for Modelling, Control and
[60] Theodore B Trafalis and Huseyin Ince. 2000. Support vector machine for re- Automation and International Conference on Intelligent Agents, Web Technologies
gression and applications to financial forecasting. In Proceedings of the IEEE- and Internet Commerce (CIMCA-IAWTIC’06), Vol. 2. IEEE, 835–841.
INNS-ENNS International Joint Conference on Neural Networks. IJCNN 2000. Neu- [69] Seongjun Yun, Minbyul Jeong, Raehyun Kim, Jaewoo Kang, and Hyunwoo J Kim.
ral Computing: New Challenges and Perspectives for the New Millennium, Vol. 6. 2019. Graph transformer networks. Advances in neural information processing
IEEE, 348–353. systems 32 (2019).
[61] Chih F Tsai and Sammy P Wang. 2009. Stock price forecasting by hybrid ma- [70] Jiawei Zhang, Haopeng Zhang, Congying Xia, and Li Sun. 2020. Graph-bert:
chine learning techniques. In Proceedings of the international multiconference of Only attention is needed for learning graph representations. arXiv preprint
engineers and computer scientists, Vol. 1. 60. arXiv:2001.05140 (2020).
[62] Manuel R Vargas, Beatriz SLP De Lima, and Alexandre G Evsukoff. 2017. Deep [71] Zhuosheng Zhang, Aston Zhang, Mu Li, Hai Zhao, George Karypis, and Alex
learning for stock market prediction from financial news articles. In 2017 IEEE Smola. 2023. Multimodal chain-of-thought reasoning in language models. arXiv
international conference on computational intelligence and virtual environments preprint arXiv:2302.00923 (2023).
for measurement systems and applications (CIVEMSA). IEEE, 60–65. [72] Wayne Xin Zhao, Kun Zhou, Junyi Li, Tianyi Tang, Xiaolei Wang, Yupeng Hou,
[63] Jar-Long Wang and Shu-Hui Chan. 2006. Stock market trading rule discovery Yingqian Min, Beichen Zhang, Junjie Zhang, Zican Dong, et al. 2023. A survey
using two-layer bias decision tree. Expert Systems with Applications 30, 4 (2006), of large language models. arXiv preprint arXiv:2303.18223 (2023).
605–611. [73] Eric Zivot and Jiahui Wang. 2006. Vector autoregressive models for multivariate
[64] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Ed Chi, Quoc Le, time series. Modeling financial time series with S-PLUS® (2006), 385–429.
and Denny Zhou. 2022. Chain of thought prompting elicits reasoning in large
language models. arXiv preprint arXiv:2201.11903 (2022). Received 20 February 2007; revised 12 March 2009; accepted 5 June 2009
[65] Shijie Wu, Ozan Irsoy, Steven Lu, Vadim Dabravolski, Mark Dredze, Sebas-
tian Gehrmann, Prabhanjan Kambadur, David Rosenberg, and Gideon Mann.
This figure "acm-jdslogo.png" is available in "png" format from:

http://arxiv.org/ps/2306.11025v1
This figure "knowledge_graph.png" is available in "png" format from:

http://arxiv.org/ps/2306.11025v1
This figure "workflow.png" is available in "png" format from:

http://arxiv.org/ps/2306.11025v1