redherring1
New member
- Dec
- 18
- 9
Global Markets
In this post I would like to share some automation and AI examples I have either directly coded/used or seen employed in my financial work environment (Hedge Funds & Banks). I think posting under Hedge Funds is most appropriate since majority of the items in the list would be interesting to the hedge fund crowd. (Although a few items are related to fintech and recruitment as well)
1- Delta Hedging for Derivatives Market Making at High Volume
This was implemented by a quant colleague and made another well-connected colleague's position obselete within the next 24 hours. My colleague's job was to manually enter and exit hedge positions based on our changing strategy and trading book.
If you are familiar with options or derivatives in general, when you enter a derivatives position as a bank (or hedge fund) you immediately create exposure as various greeks (delta, gamma, theta, omega etc.). These derivative greeks entail exposures to different aspects of derivatives such as underlying asset, time to maturity, interest rates and volatility.
Delta is the most commonly greek that needs hedging as it signifies exposure to underlying asset. Here is a simplistic example on a bank selling call options on BP.
Example (call option): In the context of the bank selling 10 call options on BP, the bank would likely engage in delta hedging to offset the potential loss from the call options being exercised. This would involve the bank taking a long position in BP stock or BP stock futures, as the price of these assets is likely to move in a similar direction to the price of the call options. By doing this, the bank can offset any potential loss from the call options, and limit its overall exposure to the price movements of BP stock.
Example (put option):
If the bank was selling 10 put options then its traders could delta hedge by short selling actual BP shares in the market.
Complexity
Now in theory delta hedging sounds rather simple but in practice greek hedging becomes quite messy. Firstly, hedging needs change continuously as the underlying price and option prices change. Even if the underlying asset's price was to remain unchanged for long periods of time, option loses its time value continously hence causing changes in its price. There are also other sensitivies such as gamma (measures delta's rate of change), vega (sensitivity to volatility), theta (sensitivity to intrinsic time value), rho (interest rate sensitivity). Hedging each aspect affects the others and must be maintained continously in a large trading book.
Volume
On a regular day, we would issue hundreds of derivative products with combined notional trading volume of around $500 million per day (if average leverage of these products is 10x, you can imagine $50 million actually being involved in derivatives trades.) We managed this operation
We were using 10s of sophisticated (but manual) tools and computer scripts to monitor and manage hedging of derivative trades throughout the day and in my team of 4, 1 colleague was specialized in hedging only meaning he focused solely on optimization of hedging for years. One week a quant from another department communicated with us that he can implement a model that can completely automate the hedging operations within satisfactory boundaries based on our requirements. He asked the execs for 1-week experimental modeling and testing and he nailed it after the 1-week period shortly after our hedging expert had absolutely no business. Sad but true story.
2- Automated Volatility Arbitrage
This is something I coded personally for one of the bulge bracket banks. Mentioned bank had a big arbitrage department which compared implied volatility components of liquid equivelant financial derivatives in the market. Every now and then traders cause small deviations in option prices due to different forecasts or sometimes even calculation mistakes (small or big). This bank was using a primitive tool to reverse engineer Black & Scholes formula for each derivative product being traded in the market and find their implied volatility. After which if the volatility seems high or low traders would sell or buy the options or warrants at stake.
When I was assigned with this task I knew instantly it was being done unnecessarily primitively. One problem was while the traders were calculating the implied volatility manually, prices would often move making it very difficult to identify or act on arbitrage opportunities.
So, I created a script that would simultaneously scan hundreds of financial products in the market using data feed from Bloomberg Terminal and Reuters Eikon, and give red or green alerts when there are significant differences in the volatility calculations. This caused great hype in the arbitrage department and we ended up making millions of extra profit in the first day. Long after I moved on to bigger roles, the bank continued using the script and someone actually improved it with adding threading modules to the code so that it can parallellize the processors for performance improvement and analysis of larger market data.
3- AI Banking Assistant
This is also a project I directly managed and implemented. I have collaborated with an Asian Fintech company to deploy a specific AI model that animates a human avatar based on real human bankers. This AI prototype was deployed at one of the two biggest credit card companies' AI research lab at its NYC headquarters.
AI was modeled after human bankers and capable of interacting with clients based on their retail or commercial banking needs. I found this project incredibly exciting and couldn't help but have lots of thoughts/concerns on how it might affect the finance industry within the next decade.
AI banker could be used through hologram technology as well as large immersive screens in a room. The experience especially through the life-size hologram is something to behold as you interact with the AI that looks and sounds extremely close to a sophisticated human and discusses various financial solutions with you based on your needs and queries. Now that AI is becoming more common and hologram projector technology is getting cheaper I expect the industry to reach a tipping point after which we can see snowball affect and exponential growth of these solutions.
4- Energy Consumption Machine Learning Model
This one was created by the quants in another department at the Hedge Fund I was working. We had semi-automated energy consumption models for each country in Europe. I say semi-automated because humans would enter consumption assumptions for each period (hour, day, week or month) depending on various factors such as social event, holidays, weather events, temperatures, sports events, political events etc.
When the AI models were introduced to the traders workflow, they weren't too happy. Models didn't do too well initially and they required some supervision during the training period. After around the 1-year mark, the models started performing exceptionally well and beating all the traders' forecasts in nearly all situations. This significantly lowered the traders' workload from repetitive tasks and freed up resources for more creative and imaginative alpha generating ideas. The main problem with such AI models is you don't want to rely on them too heavily as they might still make serious mistakes in outlier events (something they've never been trained with).
5- Countless Reporting Automation
I have also coded various reporting tools to automate big banks' tedious reporting tasks in the very early days of my finance career. This would count as automation although not AI. Nation-wide publishing of macroeconomic reports & trading reports, their generation from bits and pieces of market data, complimenting data with commentary templates, portfolio numbers and ratios, bank subsidiary coordination reports (very tedious, repetitive and annoying), compliance reports, sending reports through email servers, publishing at ftp servers/bank websites are some of the tasks I've automated related to financial reporting.
6- Social Media and Search Engine Data
Someone I know had success with deploying AI bots to analyze fragmented public consumer data such as store busy hours to calculate average consumer demand for specific brands and trading their equities in Long/Short Strategies. If executed masterfully, Long/Short strategies can be lucrative with low risk levels.
7- HR Screening
I haven't done this directly but I have worked with people who do it. Global financial institutions such as JP Morgan, HSBC, BofA, Morgan Stanley receive hundreds even thousands of applications for every position they post. Usually, the more junior the position the higher the demand will be due to generalized nature of the position (i.e. Graduate Trainee, Analyst, Junior Associate, General Application). Constantly going through mountains of CVs is a repetitive and mundane tasks especially at the early stages of the recruitment processes (screening and filtering applicants).
So, I know a number of companies and their HR departments use AI software to eliminate unwanted applicants' applications before they even reach their inbox/desk. These AI bots usually have a test period to compare the results with what the human resources' decision would be in each specific case and it must be at a very satisfactory level before the tools being deployed. That being said I think there is something inhumane at fully delegating the delicate first exposure of a potential candidate to robots. I can't help but think a few very special candidates being removed from the applicant pool because their cover letter is not fully understood by the AI model based on the criteria it was trained on. I might be wrong though as job application process can be manipulated by some candidates as well and emotions in general cause humans lots of trouble in decision making.
1- Delta Hedging for Derivatives Market Making at High Volume
This was implemented by a quant colleague and made another well-connected colleague's position obselete within the next 24 hours. My colleague's job was to manually enter and exit hedge positions based on our changing strategy and trading book.
If you are familiar with options or derivatives in general, when you enter a derivatives position as a bank (or hedge fund) you immediately create exposure as various greeks (delta, gamma, theta, omega etc.). These derivative greeks entail exposures to different aspects of derivatives such as underlying asset, time to maturity, interest rates and volatility.
Delta is the most commonly greek that needs hedging as it signifies exposure to underlying asset. Here is a simplistic example on a bank selling call options on BP.
Example (call option): In the context of the bank selling 10 call options on BP, the bank would likely engage in delta hedging to offset the potential loss from the call options being exercised. This would involve the bank taking a long position in BP stock or BP stock futures, as the price of these assets is likely to move in a similar direction to the price of the call options. By doing this, the bank can offset any potential loss from the call options, and limit its overall exposure to the price movements of BP stock.
Example (put option):
If the bank was selling 10 put options then its traders could delta hedge by short selling actual BP shares in the market.
Complexity
Now in theory delta hedging sounds rather simple but in practice greek hedging becomes quite messy. Firstly, hedging needs change continuously as the underlying price and option prices change. Even if the underlying asset's price was to remain unchanged for long periods of time, option loses its time value continously hence causing changes in its price. There are also other sensitivies such as gamma (measures delta's rate of change), vega (sensitivity to volatility), theta (sensitivity to intrinsic time value), rho (interest rate sensitivity). Hedging each aspect affects the others and must be maintained continously in a large trading book.
Volume
On a regular day, we would issue hundreds of derivative products with combined notional trading volume of around $500 million per day (if average leverage of these products is 10x, you can imagine $50 million actually being involved in derivatives trades.) We managed this operation
We were using 10s of sophisticated (but manual) tools and computer scripts to monitor and manage hedging of derivative trades throughout the day and in my team of 4, 1 colleague was specialized in hedging only meaning he focused solely on optimization of hedging for years. One week a quant from another department communicated with us that he can implement a model that can completely automate the hedging operations within satisfactory boundaries based on our requirements. He asked the execs for 1-week experimental modeling and testing and he nailed it after the 1-week period shortly after our hedging expert had absolutely no business. Sad but true story.
2- Automated Volatility Arbitrage
This is something I coded personally for one of the bulge bracket banks. Mentioned bank had a big arbitrage department which compared implied volatility components of liquid equivelant financial derivatives in the market. Every now and then traders cause small deviations in option prices due to different forecasts or sometimes even calculation mistakes (small or big). This bank was using a primitive tool to reverse engineer Black & Scholes formula for each derivative product being traded in the market and find their implied volatility. After which if the volatility seems high or low traders would sell or buy the options or warrants at stake.
When I was assigned with this task I knew instantly it was being done unnecessarily primitively. One problem was while the traders were calculating the implied volatility manually, prices would often move making it very difficult to identify or act on arbitrage opportunities.
So, I created a script that would simultaneously scan hundreds of financial products in the market using data feed from Bloomberg Terminal and Reuters Eikon, and give red or green alerts when there are significant differences in the volatility calculations. This caused great hype in the arbitrage department and we ended up making millions of extra profit in the first day. Long after I moved on to bigger roles, the bank continued using the script and someone actually improved it with adding threading modules to the code so that it can parallellize the processors for performance improvement and analysis of larger market data.
3- AI Banking Assistant
This is also a project I directly managed and implemented. I have collaborated with an Asian Fintech company to deploy a specific AI model that animates a human avatar based on real human bankers. This AI prototype was deployed at one of the two biggest credit card companies' AI research lab at its NYC headquarters.
AI was modeled after human bankers and capable of interacting with clients based on their retail or commercial banking needs. I found this project incredibly exciting and couldn't help but have lots of thoughts/concerns on how it might affect the finance industry within the next decade.
AI banker could be used through hologram technology as well as large immersive screens in a room. The experience especially through the life-size hologram is something to behold as you interact with the AI that looks and sounds extremely close to a sophisticated human and discusses various financial solutions with you based on your needs and queries. Now that AI is becoming more common and hologram projector technology is getting cheaper I expect the industry to reach a tipping point after which we can see snowball affect and exponential growth of these solutions.
4- Energy Consumption Machine Learning Model
This one was created by the quants in another department at the Hedge Fund I was working. We had semi-automated energy consumption models for each country in Europe. I say semi-automated because humans would enter consumption assumptions for each period (hour, day, week or month) depending on various factors such as social event, holidays, weather events, temperatures, sports events, political events etc.
When the AI models were introduced to the traders workflow, they weren't too happy. Models didn't do too well initially and they required some supervision during the training period. After around the 1-year mark, the models started performing exceptionally well and beating all the traders' forecasts in nearly all situations. This significantly lowered the traders' workload from repetitive tasks and freed up resources for more creative and imaginative alpha generating ideas. The main problem with such AI models is you don't want to rely on them too heavily as they might still make serious mistakes in outlier events (something they've never been trained with).
5- Countless Reporting Automation
I have also coded various reporting tools to automate big banks' tedious reporting tasks in the very early days of my finance career. This would count as automation although not AI. Nation-wide publishing of macroeconomic reports & trading reports, their generation from bits and pieces of market data, complimenting data with commentary templates, portfolio numbers and ratios, bank subsidiary coordination reports (very tedious, repetitive and annoying), compliance reports, sending reports through email servers, publishing at ftp servers/bank websites are some of the tasks I've automated related to financial reporting.
6- Social Media and Search Engine Data
Someone I know had success with deploying AI bots to analyze fragmented public consumer data such as store busy hours to calculate average consumer demand for specific brands and trading their equities in Long/Short Strategies. If executed masterfully, Long/Short strategies can be lucrative with low risk levels.
7- HR Screening
I haven't done this directly but I have worked with people who do it. Global financial institutions such as JP Morgan, HSBC, BofA, Morgan Stanley receive hundreds even thousands of applications for every position they post. Usually, the more junior the position the higher the demand will be due to generalized nature of the position (i.e. Graduate Trainee, Analyst, Junior Associate, General Application). Constantly going through mountains of CVs is a repetitive and mundane tasks especially at the early stages of the recruitment processes (screening and filtering applicants).
So, I know a number of companies and their HR departments use AI software to eliminate unwanted applicants' applications before they even reach their inbox/desk. These AI bots usually have a test period to compare the results with what the human resources' decision would be in each specific case and it must be at a very satisfactory level before the tools being deployed. That being said I think there is something inhumane at fully delegating the delicate first exposure of a potential candidate to robots. I can't help but think a few very special candidates being removed from the applicant pool because their cover letter is not fully understood by the AI model based on the criteria it was trained on. I might be wrong though as job application process can be manipulated by some candidates as well and emotions in general cause humans lots of trouble in decision making.
