Short-term stock price volatility predicting method based on the transcript of earnings conference call

Abstract

This invention discloses a method for predicting short-term stock price volatility based on transcripts of earnings conferences. The method includes the following steps: obtaining multiple transcripts of earnings conferences, extracting multiple financial feature vectors, inputting them into a two-layer neural network, and training the network using an additive-delete learning algorithm. A first weight adjustment procedure and a sample selection procedure are executed to select a subset of training samples. A second weight adjustment procedure is executed to determine if the network is acceptable. A hidden node expansion procedure is executed to add multiple hidden nodes, and the network is then judged to be acceptable. A hidden node reduction procedure is executed to select and discard hidden nodes, and the network is then judged to be acceptable. A final judgment procedure is used to determine whether the second learning objective has been achieved. If yes, the prediction result is output; otherwise, samples are reselected for the learning and training process.

Claims

1. A method for predicting short-term stock price volatility based on transcripts of investor conferences, comprising the following steps: obtaining a plurality of transcripts of investor conferences; processing the plurality of transcripts using a large language model to segment a plurality of financial information segments; and extracting features from the plurality of financial information segments to obtain a plurality of financial feature vectors; inputting a plurality of learning samples formed by the plurality of financial feature vectors into a two-layer neural network, the two-layer neural network comprising an input layer, a hidden layer, and an output layer, wherein a plurality of input nodes of the input layer receive the plurality of financial feature vectors from the plurality of learning samples; performing a learning and training procedure for the plurality of financial feature vectors using an additive learning algorithm, setting an initial hidden node and an initial weight, and performing a first weight adjustment procedure on the plurality of learning samples to adjust the initial weight to a first weight; A sample selection procedure is executed to select a subset of training samples. These samples are sorted according to their computational errors, and samples with computational errors less than a first learning objective and the first sample with a learning error greater than the first learning objective are selected as the training sample subset. A second weight adjustment procedure is executed to determine whether the computational error of the training sample subset reaches the first learning objective. If yes, the two-layer neural network is deemed acceptable; otherwise, it is deemed unacceptable. The first weight is adjusted to a second weight, and a hidden node expansion procedure is executed. The hidden node expansion procedure adds a plurality of hidden nodes. The second weight adjustment procedure is executed again to determine whether the computational error of the training sample subset reaches the first learning objective. If yes, the two-layer neural network is deemed acceptable; otherwise, the training sample subset is discarded, and the process proceeds to a final judgment procedure. When the two-layer neural network is an acceptable network, a hidden node reduction procedure is executed. A hidden node is selected to be discarded, and it is determined whether the two-layer neural network with the discarded hidden node is still acceptable. If yes, the new architecture two-layer neural network is stored, and the process proceeds to the final judgment procedure. If no, the original architecture two-layer neural network is stored again, and the process proceeds to the final judgment procedure. The final judgment procedure determines whether the proportion of training data in the training sample subset that cannot be learned reaches a second learning objective. If yes, data training is completed, and a prediction result is output by an output node of the output layer. If no, samples are reselected for the learning training procedure. The hidden node reduction procedure includes determining whether the number of discarded hidden nodes exceeds the number of hidden nodes. If yes, the two-layer neural network is considered acceptable. If no, a normalization procedure is executed, including reverse discarding, determining whether the first learning objective is met, backward operation, and weight adjustment. The second weight adjustment procedure is then re-executed until the number of discarded hidden nodes exceeds the number of hidden nodes.

2. The method for predicting short-term stock price volatility based on the verbatim transcript of earnings calls as described in claim 1, wherein the large language model includes retrieval enhancement generation technology and performs feature extraction through the FinBERT language model, and the plurality of financial information includes financial indicators, cost aspects, employee and manager aspects, business development aspects, future development aspects, and expansion aspects.

3. The short-term stock price volatility prediction method based on the verbatim transcript of the earnings call, as described in Request 1, wherein the prediction result includes a three-day stock price volatility indicator or a seven-day stock price volatility indicator.

4. The short-term stock price volatility prediction method based on the verbatim transcript of the earnings conference as described in claim 1, wherein the first weight adjustment procedure includes performing forward calculations, backward calculations and weight adjustments on the plurality of learning samples, and the first weight adjustment procedure adjusts the learning rate in each calculation cycle until the total number of cycles is reached.

5. The short-term stock price volatility prediction method based on the verbatim transcript of the earnings conference as described in Request 1, wherein the calculation error is the difference between the actual output value and the expected output value of the learning sample.

6. The short-term stock price volatility prediction method based on the verbatim transcript of the earnings conference as described in claim 1, wherein the second weight adjustment procedure includes performing forward operation on the training sample subset, determining whether the first learning objective is met, backward operation, and weight adjustment, wherein the second weight adjustment procedure adjusts the learning rate in each operation cycle until the total number of cycles is reached.

7. The short-term stock price volatility prediction method based on the earnings call transcript as described in claim 1, wherein the first learning objective of the three-day stock price volatility indicator is 0.75 and the first learning objective of the seven-day stock price volatility indicator is 0.

5.

8. The method for predicting short-term stock price volatility based on the earnings call transcript as described in claim 1, wherein the hidden node expansion procedure includes using an isolation module and adding three new hidden nodes and assigning relevant weights.

9. The short-term stock price volatility prediction method based on the earnings call transcript as described in claim 1, wherein the second learning objective of the three-day stock price volatility indicator is 0.3 and the second learning objective of the seven-day stock price volatility indicator is 0.2.

Images