Towards Earnings Call and Stock Price Movement
TTowards Earnings Call and Stock Price Movement
Zhiqiang Ma
S&P GlobalNew York, NY, [email protected]
Grace Bang
S&P GlobalNew York, NY, [email protected]
Chong Wang
S&P GlobalNew York, NY, [email protected]
Xiaomo Liu
S&P GlobalNew York, NY, [email protected]
ABSTRACT
Earnings calls are hosted by management of public companies todiscuss the company’s financial performance with analysts andinvestors. Information disclosed during an earnings call is an essen-tial source of data for analysts and investors to make investmentdecisions. Thus, we leverage earnings call transcripts to predictfuture stock price dynamics. We propose to model the language intranscripts using a deep learning framework, where an attentionmechanism is applied to encode the text data into vectors for thediscriminative network classifier to predict stock price movements.Our empirical experiments show that the proposed model is supe-rior to the traditional machine learning baselines and earnings callinformation can boost the stock price prediction performance.
CCS CONCEPTS • Computing methodologies → Supervised learning ; Naturallanguage processing . KEYWORDS stock price movement prediction; earnings call; deep learning
ACM Reference Format:
Zhiqiang Ma, Grace Bang, Chong Wang, and Xiaomo Liu. 2020. TowardsEarnings Call and Stock Price Movement. In
KDD Workshop on MachineLearning in Finance (MLF ’20), August 24, 2020, Virtual Event, USA.
ACM,New York, NY, USA, 5 pages.
With $74 trillion in assets under management in the US alone ,understanding the mechanism of stock market movements is ofgreat interest to financial analysts and researchers. As such, therehas been significant research in modeling stock market movementsusing statistical and, more recently, machine learning models inthe past few decades. However, it may not be sensible to directlypredict future stock prices given the possibility that they follow arandom walk [13]. Thus researchers have proposed to predict thedirectional movements of stocks and their volatility levels [9, 22, 26].In this study, we explore company’s earnings call transcripts dataand investigate using the information embedded in earnings calltranscripts to address the task of predicting the movements ofstocks by leveraging the recent advancements in natural languageprocessing (NLP).Stock markets demonstrate notably higher levels of volatility,trading volume, and spreads prior to earnings announcements giventhe uncertainty in company performance [7]. Such movements canbe costly to the investors as they can result in higher trading fees,missed buying opportunities, or overall position losses. Thus, the ability to accurately identify directional movements in stock pricesand hold positions accordingly based on earnings releases can behugely beneficial to investors by potentially minimizing their lossesand generating higher returns on invested assets.Stock market prices are driven by a number of factors includingnews, market sentiment, and company financial performance. Pre-dicting stock price movements based on market sentiment from thenews and social media have been studied previously [6, 9, 26]. How-ever, earnings calls, which occur when companies report on andexplain their financial results, have not been extensively studiedfor predicting stock price movements.Earnings call are conference calls hosted by the companies andoccur between the senior executives of publicly traded compa-nies and call participants such as investors and equity analysts.Generally, the earnings calls are comprised of two components: 1)Presentation of recent financial performance by senior companyexecutives and 2) Question and Answer (Q&A) session betweencompany management and market participants. Earnings calls arecomprised of tremendous insights regarding current operations andoutlook of companies, which could affect confidence and attitudeof investors towards companies and therefore result in stock pricemovements. The first part of the earnings call – Presentation – istypically scripted and rehearsed, particularly in the face of badnews. However, the question and answer portion of the call incor-porates unscripted and dynamic interactions between the marketparticipants and management thus allowing for a more authenticassessment of a company. Thus, we focus on the Answer section inthis work and discuss our findings regarding Presentation data inSection 6.In this paper, we propose a deep learning network to predictthe stock price movement, in which sentences from the Answersection of a transcript are represented as vectors by aggregatingword embeddings and an attention mechanism is used to capturetheir contributions to predictions. Discrete industry categories ofcompanies are also considered in the work by encoding them intolearnable vector representations. We compare the proposed methodwith several classical machine learning algorithms to assess itseffectiveness. We review several related work and present our re-searching and findings in the reset of this paper. Stock price movement predictions have traditionally been consid-ered a time series prediction problem [26]. Existing approachestackle this problem by discovering trading patterns in the histori-cal market data to predict future movements. Statisticians usuallyuse time series analysis techniques like exponential smoothing,autoregressive (AR), and autoregressive integrated moving average(ARIMA) to predict prices or price movements. Computer science a r X i v : . [ q -f i n . S T ] A ug esearchers have also showed great interest in this topic and haveapplied machine learning prediction models [17, 19] to solving thistask. Recurrent neural networks (RNN) and especially its variantssuch as LSTM [8], which were developed to process sequentialsignals, have been widely adopted to model time series stock data[3, 16]. In contrast to these statistical methods, RNN is not subjectto the stationarity requirement on the stock time series data andis able to capture the dependency of stock prices at different timeinstances.Another important research branch on this topic concentrateson leveraging external information outside of market data, e.g.,events, news, macroeconomic environment, business operations,and geopolitical status, as drivers of stock price movements. Forexample, Equifax’s stock price plummeted more than 15 percentimmediately following news reports that it had suffered a massivedata breach scandal. [5, 6] proposed to extract structured eventsfrom news and then use deep neural networks to model the impactof the events on the stock movement. Hu et al. [9] developed ahierarchical attention based neural network – HAN – studying thedependency and influence of the recent online news on stock mar-kets. As social media began reporting breaking news, researchersfound social media posts can serve as input along with historicalstock data [20, 26]. Financial filings (10-K) summarizing companies’business performance contain sentiment signals from management,which can be used to forecast stock return volatility [24]. Bag-of-words features, TFIDF and LOG1P, were adopted in their work.Researchers name this type of prediction fundamental analysis, andsince our work adopted earnings call as the input, it falls in thiscategory as well.As mentioned in the Section 1, earnings call transcripts haveunique properties and provide crucial information of companies.There is tremendous potential for exploration of this dataset aslimited previous work has studied earnings call transcripts in stockreturn volatility prediction to evaluate companies’ financial risk[22, 25]. In [22], Theil et al. introduced a neural network PRoFETto predict the stock return volatility, where it considers textual fea-tures from earnings call transcripts and financial features includingpast volatility, market volatility, book-to-market, etc. To create thetextual feature, each section (presentation, questions, and answers)is represented as a vector by applying a Bi-LSTM with attentionmechanism [1] on the tokens. Financial features pass through adeep feedforward network to calculate the financial vector. Thefinal prediction is returned by summing these two vectors andfeeding to another hidden layer.In NLP tasks, word representation is always a critical component.Pre-trained word vectors and embeddings have been widely adoptedin various state-of-the-art NLP architectures and achieved greatsuccess. The work like word2vec [14] and GloVe [18] representswords as high dimensional real-valued vectors, and their vectorarithmetic operations can reflect the semantic relationship of thewords. In this work, we adopted the pre-trained GloVe embeddingsto save computing time. Assuming that there is a set of stocks Θ = { S , S , · · · , S n } of n public companies. For a stock S c , there exists a series of earnings call transcript Γ c = { T d , T d , · · · , T d m } , which were held on dates d , d , · · · , d m respectively. The goal is to predict the movementof the stock S c on date d i + T d i occurredon date d i . The movement y is a binary value, 0 (down) or 1 (up).The stock price in the market moves constantly in a trading day. Toformally define y , here we adopt the closing price, i.e. y = ( p d i + > p d i ) , where p d i and p d i + are the closing prices on date d i and d i + f , which takes features E extracted from an earnings call transcript T and industry cate-gorization I of the company as input, to predict the stock pricemovement y of the day after the earnings call. To solve the problem, we utilize two features to build the predic-tion function: 1) Answer section textual feature and 2) companyindustry type feature. In this section, we firstly propose a deepneural network structure designed to represent the textual feature.Hereafter, we introduce the industry type embedding. The finalprediction is generated via a discriminative network by feeding inthe combined features.
A Q&A section consists of multiple rounds of communications be-tween market participants and company management executives.We only use Answer sections from managements with the assump-tion that the answers are a more realistic representation of thefeedback interested by investors. In the case where a response pro-vided by managements does not answer a specific question, marketparticipants typically follow up with clarifying questions to whichthey then receive required answers.
Sentence Embedding:
Given an earnings call transcript T , we extractthe answer sequence A = [ l , l , · · · , l N ] and A ∈ T , l i denoting asentence that comes from splitting the Answer section. We treatone sentence as a feature atom, and transform each sentence toa dense vector. To achieve that, we process each token o of a sen-tence l to a distributed representation vector e o by leveraging apre-trained embedding layer. The sentence vector v l is constructedby concatenating two vectors obtained from average pooling andmax pooling the token vectors across all the tokens of the sentence.To reduce computing complexity, we didn’t allow the word embed-ding layer to be trainable or fine-tuned. Another popular approachto representing sentences is to employ RNN to encode a wholesentence to a hidden state vector from the last recurrent unit [22].Sentence encoders [2, 4] may be used here too. We leave them forthe future exploration. Sentence Attention:
Undoubtedly, some sentences convey more in-formation while others do not for the task of predicting stock pricemovements. We leverage the idea of the attention mechanism intro-duced in the machine translation domain [1] to learn the weightsof the sentences, where the weights quantify the contributions ofthe sentences to the final outcome. Given an answer sequence A consisting of N sentences and sentences transformed to embeddingvector v s, the attention weights α ∈ R × N are defined as normal-ized scores over all the sentences by a softmax function as shownelow, α l = softmax ( score ( v l )) , score ( v l ) = u T v l + b , (1)where u is a learnable vector parameter and b is a learnable biasparameter. The score function may be replaced with others de-pending on the specific task. Refer to [1, 12, 23] for other scorefunction options. By aggregating the sentence vectors weighted onthe attention parameter, the earnings call answer sequence can betransformed to E = N (cid:213) l α l v l . (2)Figure 1 demonstrates our network structure introduced above. Company stock prices usually follow the trend of the industry sec-tor in which it belongs. The sector category and company sectordefinition vary in terms of standards. We select the Global IndustryClassification Standard (GICS) definition in our study. GICS con-sists of 11 industry sector categories, such as ‘energy’, ‘financials’,and ‘health care’. The industry sector is a categorical indicator. Inmachine learning, categorical data are usually transformed by one-hot encoding or ordinal encoding, while we create an embeddinglayer to transform the categorical values into vector presentation I ,which is learnable during the network training phase. With the feature representations E and I built above as input, thefinal binary classification result is computed by a discriminativenetwork. The feed forward discriminative network consists of mul-tiple hidden layers — batch normalization layer [10], dropout layer[21], ReLU activation layer [15], and linear layer. Figure 2 illustratesthe complete neural network structure including the discriminativenetwork. We perform experiments by using earnings calls transcripts ofS&P 500 companies. We collected 17025 earnings call transcriptsover 485 companies from S&P Global Market Intelligence TRAN-SCRIPTS database. The temporal span of the data is between 2009and 2019, and the temporal spans for a few companies might beshorter because the companies were added to S&P 500 later than2009. On average, each company has around 35 transcripts. Everytranscript in the TRANSCRIPTS database has been segmented intocomponents in terms of types of the components, such as ‘Presen-tation Operator Message’, ‘Presentation Section’, ‘Question’, and‘Answer’. We select the ‘Answer’ components and employed NLTKsentence tokenizer to split sections to sentences. Figure 3 showsthe distribution of the number of sentences in ‘Answer’ compo-nents in the dataset. Table 1 shows more statistics of the dataset interms of sentences and tokens (stop words excluded). As to down-load the corresponding historical stock data to get the stock price S&P 500 index composes 505 stocks from 500 companies. Due to merger and acquisi-tion, ticker changing, shortness of available transcripts, and etc. reasons, 15 companieswere not included in the data. token token token token token token sentence 1 sentence N embedding layer v attention layer E xx + v N max average Figure 1: Neural network structure for learning textual fea-ture vectors. The input is tokens from sentences of an An-swer section, and the output E is a vector representation ofthe input. token token token sentence emb., average, max layersattention layerdiscriminative networkembedding layerindustry id Figure 2: Proposed neural network structure. The input tothe discriminative network is a concatenated vector of a tex-tual feature vector and an industry embedding vector.
Answer Section No. of Sentences No. of TokensTotal 3.6 M M Average 2145 1457
Table 1: Statistics of the raw text data (without stop words). movements y , we map company names to their stock tickers andemployed Python pandas_datareader package with the sourceset to yahoo . For the experiments, we hold outthe most recent five earnings call transcripts from each company igure 3: Histogram of the number of sentences in Answersections. The dashed line in red at 300 at x − axis indicatesthe cut-off on the length of the Answer section. as the testing dataset (2425 observations in total), and everythingelse is used as the training and validation dataset. Please note thatcompanies have their earnings call conferences on different datesfor every reporting quarter, and it is not reasonable to set a cutoffdate to split the dataset.We split Answer section to sentences and then tokenize sen-tences to tokens. When transforming tokens to embedding vectors,a vocabulary is constructed, where stop words are ignored and to-kens with total frequency less than four times would be disregardedas well. Tokens are transformed to vectors by applying pre-trainedGloVe (embedding dimension = E of each transcript to N = In order to assess the performance of our model,we compare its performance with two baseline models below: • Mean Reversion (MR) : MR is a simple trading strategy, whichassumes that stock prices would tend to revert toward theirmoving averages when deviating from them. We calculate60-day moving average in the experiment. • XGBoost : XGBoost has achieved great success in solvingvarious classification problems in practice. To transformthe text data into numeric format, we adopted two featureengineering techniques,
TFIDF and
LOG1P defined as below[11]: – TFIDF = TF ( o , A ) × IDF ( o , A ) = TC ( o , A ) × IDF ( o , A ) – LOG1P = LOG ( + TC ( o , A )) where TC ( o , A )) is the count of the token o in earnings callAnswer section A and IDF ( o , A ) = log (| Γ |/| A ∈ Γ , o ∈ A |) . In the experiments, we predict the stock price move-ment of the companies on the day just after their earnings callbeing released. Table 2 compares the performance of our proposedmodel and the baseline models. When compare the models, weadopt two evaluation metrics, accuracy and Matthews CorrelationCoefficient (MCC), which are also used in the previous work [6, 26].
Table 2: Model performance summary.
Accuracy (%) MCCMR 50.80 0.0202XGBoost (Log1P) 50.89 0.0013XGBoost (TFIDF) 51.25 0.0154Our Model
The definition of MCC is as follows, given true positive (tp), truenegative (tn), false positive (fp), and false negative (fn) from theprediction output:MCC = tp · tn − fp · fn (cid:112) ( tp + fp )( tp + fn )( tn + fp )( tn + fn ) . The value of MCC is between − − Undoubtedly, stock price movement prediction is a very challeng-ing task. Through our experimentation in this study, we confirmthat earnings call transcripts have certain predictive power for fu-ture stock price movements. Thus, the inclusion of this dataset instock price prediction analysis can have predictive impact in theevelopment of such systems for in practice use of stock investmentrisk analysis. Additionally, we note two other aspects, which areworth more investigation in the future: • In Section 1, we mention that only Answer sections areincluded in the model with the management Presentationsections excluded. Our decision on that, besides the heuris-tics reason mentioned in Section 1, is that the model doesnot improve by including the Presentation data. The soleconsideration of the Presentation text also did not improvethe model performance results. Interestingly, this observa-tion is not consistent with the conclusion made by Theil etal. [22], where the Presentation data yields better results intheir ablation study for predicting stock volatility, despitethe different prediction targets in their work and ours. Futurework will be conducted to justify the observation. • In addition to the fundamental analysis like our work, fea-tures originating from technical analysis on the historicalstock price data are able to be absorbed into the forecastmodel. For example, the historical stock time series data canbe encoded into another feature vectors by RNN models,which are further used to build global vectors along with thefeatures from the fundamental analysis.To summarize, we propose leveraging textual information fromAnswer sections of earnings call transcripts to predict movementsof stock prices. To create textual features from transcripts, tokensare transformed into embedding vectors and sentence vectors arebuilt by max pooling and average pooling over the word vectors.An earnings call Answer section is represented as a vector by aggre-gating its sentence vectors through the attention mechanism. Thefinal prediction is made by a discriminative network which takesthe textual feature vectors and learned industry embedding vectorsas input. The experiments show that the proposed deep learningmodel outperforms the classical baseline models, and also provethat the information conveyed in earning calls correlates with stockprice movements and therefore can be used in relevant forecastingtasks.
REFERENCES [1] Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. 2015. Neural MachineTranslation by Jointly Learning to Align and Translate. In .[2] Daniel Cer, Yinfei Yang, Sheng-yi Kong, Nan Hua, Nicole Limtiaco, RhomniSt. John, Noah Constant, Mario Guajardo-Cespedes, Steve Yuan, Chris Tar, BrianStrope, and Ray Kurzweil. 2018. Universal Sentence Encoder for English. In
Proceedings of the 2018 Conference on Empirical Methods in Natural LanguageProcessing: System Demonstrations . Brussels, Belgium, 169–174.[3] L. Cheng, Y. Huang, and M. Wu. 2018. Applied attention-based LSTM neuralnetworks in stock prediction. In . 4716–4718.[4] Alexis Conneau, Douwe Kiela, Holger Schwenk, Loïc Barrault, and AntoineBordes. 2017. Supervised Learning of Universal Sentence Representations fromNatural Language Inference Data. In
Proceedings of the 2017 Conference on Em-pirical Methods in Natural Language Processing . Association for ComputationalLinguistics, Copenhagen, Denmark, 670–680.[5] Shumin Deng, Ningyu Zhang, Wen Zhang, Jiaoyan Chen, Jeff Z. Pan, and HuajunChen. 2019. Knowledge-Driven Stock Trend Prediction and Explanation viaTemporal Convolutional Network. In
Companion Proceedings of The 2019 WorldWide Web Conference (WWW ’19) . Association for Computing Machinery, NewYork, NY, USA, 678–685.[6] Xiao Ding, Yue Zhang, Ting Liu, and Junwen Duan. 2015. Deep Learning forEvent-Driven Stock Prediction. In
Proceedings of the 24th International Conferenceon Artificial Intelligence (IJCAI ’15) . AAAI Press, 2327–2333. [7] W.M. Donders, Monique, Roy Kouwenberg, and C. F. Vorst, Ton. 2000. Optionsand earnings announcements: an empirical study of volatility, trading volume,open interest and liquidity.
European Financial Management
6, 2 (2000), 149–171.[8] Sepp Hochreiter and Jürgen Schmidhuber. 1997. Long Short-Term Memory.
Neural Computation
9, 8 (1997), 1735–1780.[9] Ziniu Hu, Weiqing Liu, Jiang Bian, Xuanzhe Liu, and Tie-Yan Liu. 2018. Listeningto Chaotic Whispers: A Deep Learning Framework for News-Oriented StockTrend Prediction. In
Proceedings of the Eleventh ACM International Conference onWeb Search and Data Mining (WSDM ’18) . Association for Computing Machinery,New York, NY, USA, 261–269.[10] Sergey Ioffe and Christian Szegedy. 2015. Batch Normalization: AcceleratingDeep Network Training by Reducing Internal Covariate Shift. In
Proceedings ofthe 32nd International Conference on International Conference on Machine Learning- Volume 37 (ICML ’15) . JMLR.org, 448–456.[11] Shimon Kogan, Dimitry Levin, Bryan R. Routledge, Jacob S. Sagi, and Noah A.Smith. 2009. Predicting Risk from Financial Reports with Regression. In
Proceed-ings of Human Language Technologies: The 2009 Annual Conference of the NorthAmerican Chapter of the Association for Computational Linguistics . Associationfor Computational Linguistics, Boulder, Colorado, 272–280.[12] Thang Luong, Hieu Pham, and Christopher D. Manning. 2015. Effective Ap-proaches to Attention-based Neural Machine Translation. In
Proceedings of the2015 Conference on Empirical Methods in Natural Language Processing . Associationfor Computational Linguistics, Lisbon, Portugal, 1412–1421.[13] Burton Gordon Malkiel. 1999.
A Random Walk down Wall Street: The Time-testedStrategy for Successful Investing . WW Norton & Company.[14] Tomas Mikolov, Kai Chen, Greg Corrado, and Jeffrey Dean. 2013. EfficientEstimation of Word Representations in Vector Space.
CoRR abs/1301.3781 (2013).[15] Vinod Nair and Geoffrey E. Hinton. 2010. Rectified Linear Units Improve Re-stricted Boltzmann Machines. In
Proceedings of the 27th International Conferenceon International Conference on Machine Learning (ICML’ 10) . Madison, WI, USA,807–814.[16] D. M. Q. Nelson, A. C. M. Pereira, and R. A. de Oliveira. 2017. Stock market’sprice movement prediction with LSTM neural networks. In . 1419–1426.[17] Jigar Patel, Sahil Shah, Priyank Thakkar, and Ketan Kotecha. 2014. Predictingstock market index using fusion of machine learning techniques.
Expert Systemswith Applications
42, 4 (10 2014), 2162–2172.[18] Jeffrey Pennington, Richard Socher, and Christopher D. Manning. 2014. GloVe:Global Vectors for Word Representation. In
Empirical Methods in Natural Lan-guage Processing (EMNLP) . 1532–1543.[19] Alaa Sheta, Sara Ahmed, and Hossam Faris. 2015. A Comparison between Re-gression, Artificial Neural Networks and Support Vector Machines for PredictingStock Market Index.
International Journal of Advanced Research in ArtificialIntelligence
Proceedings ofthe 51st Annual Meeting of the Association for Computational Linguistics (Volume 2:Short Papers) . Association for Computational Linguistics, Sofia, Bulgaria, 24–29.[21] Nitish Srivastava, Geoffrey Hinton, Alex Krizhevsky, Ilya Sutskever, and RuslanSalakhutdinov. 2014. Dropout: A Simple Way to Prevent Neural Networks fromOverfitting.
J. Mach. Learn. Res.
15, 1 (Jan. 2014), 1929–1958.[22] Christoph Kilian Theil, Samuel Broscheit, and Heiner Stuckenschmidt. 2019.PRoFET: Predicting the Risk of Firms from Event Transcripts. In
Proceedings of theTwenty-Eighth International Joint Conference on Artificial Intelligence (IJCAI ’19) .International Joint Conferences on Artificial Intelligence Organization, 5211–5217.[23] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones,Aidan N. Gomez, undefinedukasz Kaiser, and Illia Polosukhin. 2017. Atten-tion is All You Need. In
Proceedings of the 31st International Conference on NeuralInformation Processing Systems (NIPS’ 17) . Curran Associates Inc., Red Hook, NY,USA, 6000–6010.[24] Chuan-Ju Wang, Ming-Feng Tsai, Tse Liu, and Chin-Ting Chang. 2013. FinancialSentiment Analysis for Risk Prediction. In
Proceedings of the Sixth InternationalJoint Conference on Natural Language Processing . Asian Federation of NaturalLanguage Processing, Nagoya, Japan, 802–808.[25] William Yang Wang and Zhenhao Hua. 2014. A Semiparametric Gaussian CopulaRegression Model for Predicting Financial Risks from Earnings Calls. In
Proceed-ings of the 52nd Annual Meeting of the Association for Computational Linguistics(Volume 1: Long Papers) . Baltimore, Maryland, 1155–1165.[26] Yumo Xu and Shay B. Cohen. 2018. Stock Movement Prediction from Tweets andHistorical Prices. In