News and Market Sentiment Analytics

class: center, middle, inverse, title-slide

.title[
# News and Market Sentiment Analytics
]
.subtitle[
## Lecture 7: RAGs
]
.author[
### Christian Vedel,<br> Department of Economics<br><br> Email: <a href="mailto:christian-vs@sam.sdu.dk" class="email">christian-vs@sam.sdu.dk</a>
]
.date[
### Updated 2025-12-12
]

---

.pull-left-wide {
  float: left;
  width: 66%;
}
.pull-right-wide {
  float: right;
  width: 66%;
}
.pull-right-wide ~ p {
  clear: both;
}

.pull-left-narrow {
  float: left;
  width: 30%;
}
.pull-right-narrow {
  float: right;
  width: 30%;
}

.small123 {
  font-size: 0.80em;
}

.large123 {
  font-size: 2em;
}

.red {
  color: red
}
</style>

# Last time
.pull-left[
**Two tedious common tasks:**
- Record Linking
- Topic Modeling 
]

.pull-right-narrow[
![Trees](Figures/Trees.jpg)
]

---
class: middle
# Today's lecture
.pull-left[
- RAGs: Back to the original purpose
- Course evaluation
]

.pull-right-narrow[
![Flows](Figures/0_Transformations, dataflows, magical, mathematical _esrgan-v1-x2plus.png)
]

---
# Motivaiton for the course

.pull-left[
- *It's all about information*
- Text is data. Data is text. 
- Applications: Business, Finance, Research, Software Development, Psychology, Linguistics, Business Intelligence, Policy Evaluation 
- In a world of LLMs you want to understand what is happening behind the fancy interface 
- The techniques live inside crazy sci-fi technologies, that you can learn. E.g. [literal mind reading](https://www.nature.com/articles/s41593-023-01304-9)
]

.pull-right[
![Counterexample](Figures/machine_learning.png)
*Counter-example to what we will be doing*
*[[xkcd.com/1838]](xkcd.com/1838)*
]

---
# Efficient Market Hypothesis (EMH)

.pull-left[
*If you can buy cheaply something which is more valuable than its price, you can profit (arbitrage). But ...*
1. We all have access to the same information 
2. Everyone is interested in maximum profits
3. Everyone buys the undervalued assets 
4. Prices increases until the advantages is gone

]
.pull-right[
![Efficient Market](Figures/stonks.jpeg)

### Implications:
- If any 'whole' in the market occurs it is closed within milliseconds 
- We cannot invest better than anyone else 
- Investing in what a chicken poops on is better than listening to an 'expert'
]

---

# Information Extraction in Financial Markets

.pull-left[
- The NLP promise is to be able to extract information 
- News about firms changes the expected returns on assets 
- [NVIDIA News](https://youtu.be/U0dHDr0WFmQ?si=EKAQgBiTekQhxl0u)
- If the expected returns on an asset increases or decreases there is an opportunity for arbitrage if prices have not changed
- NLP tools potentially allow us to be ahead 
- EMH dissallows open source implementations: You need to improve on the methods you get here
- **Timing is key**
]

.pull-right[
![Information Extraction](Figures/big_stonks.jpg)
]

---
class: middle
# RAGs
- The modern approach to sentiment analysis
- Potent context awareness
- 'simple' networks of LLMs

---
class: middle

# RAG basics
- **Query**: Question you want answered. E.g. "Is the market bullish or bearish?"
- **Augmentation**: We want to augment this query with relevant information
- **Retrieval**: We retrieve relevant documents for the query
- **Generation**: We use this information to generate an answer. E.g. "bull"
- **R  A  G**

---
class: middle

# RAG applications
- Sentiment analysis
- Systematic document review
- Legal discovery / case review
- LLM magic

---
class: middle
# Retrieval
- **Library:** Need to define and source a library. E.g. all press releases, etc.
- For any given query, we need to find a relevant match document in the library
- Approaches: Lessons from [Record Linking](https://raw.githack.com/christianvedels/News_and_Market_Sentiment_Analytics/refs/heads/main/Lecture%206%20-%20Record%20linking/Slides.html) + Searching embedding space

---
class: middle
# Retrieval
- **Query**: "How is NVIDIA doing?"
- **Library**: Press releases. Among which "NVIDIA releases devastating news"
- **Embedding approach**:
  + Embed the entire library
  + Embed the query
  + Find minimum cosine distance
- **Other approaches**:
  + Run entity recognition from e.g. spaCy
- **Trick**: Embed everything in the library beforehand and store it in a vector library

---
class: middle
# Augementation
- You simply prompt the LLM with the added documents as context.
- You can consider prompt tuning as in [arxiv.org/abs/2310.04027](https://arxiv.org/abs/2310.04027)

---
class: middle

.pull-left-wide[
# Generation
- It is hard to get LLMs to behave:
  + *"Please for the love of everything you hold dear, just say 'yes' or 'no'. Otherwise I will forever hold you in contempt and fire you in shame. If you write ONLY 'yes' or 'no', you are my favorite employee and I will give you a huge bonus"*
- (*Or restrict output space - available in some models.*)
]

---
class: middle
# A very simple RAG
### ['SDU News RAG'](https://github.com/christianvedels/News_and_Market_Sentiment_Analytics/tree/main/Lecture%207%20-%20RAGs/Code)

#### Basic structure:
- Query: Questions about the market
- Library: News headlines
- Retrieval: Base on BERT embeddings
- Generation: 
  + Expert (analysis) `\(\rightarrow\)` Censor `\(\rightarrow\)` Expert (conclusion) `\(\rightarrow\)` Zero-shot mapping to 'bear', 'bull', 'neural'
  
.footnote[
Confused about classes? Check [Code/Simple_class_example.py](https://github.com/christianvedels/News_and_Market_Sentiment_Analytics/blob/main/Lecture%207%20-%20RAGs/Code/simple_class_example.py)
]

---
class: middle
# Course evaluation

.pull-left[
- **Keep, Add, Drop (KAD)**
- [Link to online form](https://forms.gle/RRnTbEsdBzKYrtrWA)
]

.pull-right[
![KAT](Figures/KAT.png)
*A 'KAD'*
]

---
class: middle

# Exam workshop

.pull-left[
- Friday December 19th 10:15-12:00
  + **Also available:** Tuesday December 16th 10:30-12:00
- Zoom link will be available on itslearning

### Format of the workshop
- 5-10 min. presentation on what you are working on
- Room for asking questions from your peers
- **This will not impact your grade in any way**
- My feedback will be limited but I will try to facilitate useful peer feedback.
]

.pull-right[
### What you have to do
1. Have a look at the exam
2. Decide on a project you want to work on
3. Send me an email before Friday [christian-vs@sam.sdu.dk]
    + "Hi Christian, I will be working on [x]. I would like to give a 5-10 min. presentation on what I have been working on, on Monday".

- If noone is interested, we will of course just cancel.
    + No shame if it is not useful to you. Then I believe, that you know how to spend your exam time most wisely. 
]