操作步骤如下：
- 我们使用ChatGPT 为每个新闻标题提供推荐并将其转换为“ChatGPT 分数”，其中“是”映射到 1，“未知”映射到 0，“否”映射到 -1。
- 如果一家公司在某一天有多个头条新闻，我们会平均得分。
- 我们将头条新闻与下一个市场周期相匹配。 对于开盘日早上6点之前的头条，我们假设头条可以在当天开市前交易，并在当天收盘时卖出。 对于早上 6 点之后下午 4 点之前的头条，我们假设头条可以在当天收盘时交易，并在第二天收盘时卖出。 对于下午4点以后的头条，我们假设头条可以在次日的开盘价成交，在次日的收盘价卖出。
- 我们对 ChatGPT 分数的第二天回报进行线性回归，并将其与新闻策划公司提供的情绪分数进行比较。 因此，我们所有的结果都是样本外的。
原文：
We prompt ChatGPT to provide a recommendation for each headline and transform it into a “ChatGPT score,” where “YES” is mapped to 1, “UNKNOWN” to 0, and “NO” to -1. We average the scores if there are multiple headlines for a company on a given day. We match the headlines to the next market period. For headlines before 6 am on the opening day, we assume the headlines can be traded by the market opening of the same day and sold at the close of the same day. For headlines after 6 am but before 4 pm, we assume the headlines can be traded at the same day’s close and sold at the close of the next day. For headlines after 4 pm, we assume the headlines can be traded at the opening price of the next day and sold at the closing price of that next day. We then run linear regressions of the next day’s returns on the ChatGPT score and compare it to the sentiment score provided by a news curating company. Thus, all of our results are out-of-sample.
- 参考文献：
https://arxiv.org/pdf/2304.07619.pdf
