(The author of this article is Zhang Xiaoquan, Irwin and Joan Jacobs Chair Professor of School of Economics and Management, Tsinghua University)
"The biggest difference between winners and losers is that winners learn from their own and other people's mistakes."
——Sir John Templeton (famous investor)
The previous article told the story of ltcm’s prosperity and decline. In fact, this story may have another direction.
Renaissance Technologies is also composed of a group of highly intelligent mathematicians and scientists who can be called geniuses. They also rely on algorithms and models for trading. They have also achieved great success and created the myth of a money-making machine.
However, ltcm opened high and went low, becoming a footnote in history. Later people warned that renaissance is still continuing to write its own legend.
Although the two are somewhat similar, their differences are the key to success or failure. Let’s do a comparison and analysis from a quantitative perspective.
Short-term prediction is better than long-term prediction
Since algorithms and models are used for trading, some people classify ltcm into the category of quantitative trading, which is still controversial.
But from a product level, the first obvious difference between renaissance and ltcm is that they conduct different types of transactions, and the duration of the transactions is also completely different.
ltcm trades primarily in government bond pairs. These trades take several months to complete, so accurate forecasts need to be made over several months. Renaissance, on the other hand, only holds a few days (sometimes even a few hours) of trades and only requires predictions for the entire day.
All else being equal, a shorter forecast period means more efficient training data for time series modeling. This allows renaissance to quickly identify when its models are biased in the real world. One of the most dangerous ways a model can fail is if it performs well in backtests but performs poorly in real life. renaissance is able to discover within a day or two whether a model validated by backtesting fails in practice.
There were many similar failed renaissances in the early years, but because they were discovered quickly, the losses were relatively small.
In contrast, it took several months for ltcm to discover that its model performed differently in reality than in backtests. By the time they first noticed a major model failure, it was too late; the company was on the verge of bankruptcy due to months of bad exposure that had accumulated.
Restrained growth and scale control
Compared with the huge success in performance, renaissance's product line has grown very little.
renaissance started out trading commodities and currencies, and by the mid-1980s it had mastered how to create sustainable but relatively small profits. To grow, it needs to tap into a bigger market: stocks. Renaissance has been testing stock trading algorithms for nearly a decade and found many that succeeded in backtesting but failed in the real world.
Some outstanding talents have left due to the frustration caused by constant attempts and failures. Some people think that Renaissance is too cautious about stocks, while others think that the stock problem cannot be solved. Fortunately, the core team of scientists persisted, and they finally found an algorithm that works in reality. This is the result of hundreds of incremental tweaks and bug fixes, not one moment of “aha.” Renaissance has slowly expanded this stock algorithm over the years.
Even as it scales up its equities business, Renaissance has taken a balancing act to achieve modest growth. In 2003, Simons concluded that Renaissance was managing more money than its model could comfortably allocate, so the company stopped taking money from outside investors, a decision that remained unchanged for years.
In contrast, ltcm's product types can be said to be in full bloom. It initially focused on sovereign debt convergence trading, and other firms took note of its success and began copying its trades. Within 24 months of its founding, the influx of competition drove down its profits, and ltcm looked elsewhere, subsequently investing heavily in companies with dual-class structures distinct from bonds and in merger arbitrage.
In fact, ltcm failed to demonstrate sustainable profitability in its core products and moved to new and even unproven products. This move is regarded as a sign of success by the outside world and the media.
For time series forecasting products, the best way to verify that success is real (and not just luck) is to demonstrate real-world accuracy over a long period of time. renaissance does this, ltcm does not.
Never doubt a model because of your intuition, and never completely trust a model.
ltcm was founded with a focus on one product: algorithmically generated sovereign debt convergence trades. The company believes in the power of its original algorithm—it’s based on the work of two Nobel Prize winners and has an enviable track record. Unfortunately, this record was short-lived. Based on very few results (due to the long-term nature of the trade), it is indistinguishable from a lucky streak. As a result, ltcm completely believed in its model.
has expanded from convergence transactions to merger arbitrage and dual-class companies, forcing ltcm to rely more on the intuition of its partners.
Of course, ltcm built models for these new products. However, there is no model to allocate risk between different products (asset classes), this part of the work is done by committees. Risk allocation ultimately reflects committee members' beliefs and internal politics.
For example, the head of ltcm's London office had influence within the company and was passionate about a particular Royal Dutch Shell deal. Therefore, despite the high concentration of risks and lack of successful experience in this transaction, ltcm invested more than 2 billion US dollars.
Let’s look at renaissance.
Simons was frustrated many times early in his investing career by ignoring algorithmic advice, and along the way he learned that you can and should recognize when a model has serious problems, stop trading on its recommendations, and fix them. At the same time, he also realized that if he was generally uncomfortable with the model's decisions, but the model's decisions were not obviously wrong, then his intuition would not produce better results.
In renaissance, it is commonplace to turn off a model, debug it, and then turn it back on again, but suggestions to overturn a model are irreversible; any reasons for overturning a decision must either be embedded in the model itself, or be discarded.
renaissance also benefits from its single trading model, which is responsible for all decisions for the company’s flagship fund. In software development, monolithic codebase has become a pejorative term, but it brings an interesting advantage to renaissance: it forces distribution trade-offs between products to be managed algorithmically. This differs from LTCM’s hybrid human-machine model, where the model picks the assets but a committee decides how to allocate risk among different asset classes.
Differentiated Data
The final factor that separates renaissance from ltcm is renaissance's focus on differentiated data.
renaissance started out by focusing on a relatively common data type: daily closing prices for commodities and currencies. This is data that other traders will also use. But renaissance realizes that differentiated data doesn’t just mean different types of data, it also means higher quality and broader coverage of the same data type compared to competitors.
Actually, it makes a lot of sense to focus on the same data types as your competitors.
The fact that all other trading firms value this data type shows that this data type provides a strong signal. Moreover, it is often easier and more cost-effective to amplify a known valuable signal than to find a completely new signal.
renaissance amplifies the signal by gathering daily price data further back than competitors, integrating data from more sources, and cleaning and cross-checking the data to eliminate erroneous values.
As the company grew, renaissance began to collect new data types: from daily closing prices to intraday prices, to company financial data, to free-text news reports...
In hindsight, it is easy to discount the advantages of renaissanceltcm Compare the shortcomings.But renaissance is not perfect, and ltcm also has some very innovative and valuable ideas that were quietly adopted by other companies after its collapse.
The real world is complex, but there is a simple rule of thumb that applies even to startups that have nothing to do with finance or algorithms: get the core product right before scaling up. Don’t confuse accolades from the media, investors, or even customers with the ability to consistently turn a profit. Ultimately, this is what ultimately allows a business to control its own destiny.
This article represents the author's views only.
(The author of this article is Zhang Xiaoquan, Irwin and Joan Jacobs Chair Professor of School of Economics and Management, Tsinghua University)
"The biggest difference between winners and losers is that winners learn from their own and other people's mistakes."
——Sir John Templeton (famous investor)
The previous article told the story of ltcm’s prosperity and decline. In fact, this story may have another direction.
Renaissance Technologies is also composed of a group of highly intelligent mathematicians and scientists who can be called geniuses. They also rely on algorithms and models for trading. They have also achieved great success and created the myth of a money-making machine.
However, ltcm opened high and went low, becoming a footnote in history. Later people warned that renaissance is still continuing to write its own legend.
Although the two are somewhat similar, their differences are the key to success or failure. Let’s do a comparison and analysis from a quantitative perspective.
Short-term prediction is better than long-term prediction
Since algorithms and models are used for trading, some people classify ltcm into the category of quantitative trading, which is still controversial.
But from a product level, the first obvious difference between renaissance and ltcm is that they conduct different types of transactions, and the duration of the transactions is also completely different.
ltcm trades primarily in government bond pairs. These trades take several months to complete, so accurate forecasts need to be made over several months. Renaissance, on the other hand, only holds a few days (sometimes even a few hours) of trades and only requires predictions for the entire day.
All else being equal, a shorter forecast period means more efficient training data for time series modeling. This allows renaissance to quickly identify when its models are biased in the real world. One of the most dangerous ways a model can fail is if it performs well in backtests but performs poorly in real life. renaissance is able to discover within a day or two whether a model validated by backtesting fails in practice.
There were many similar failed renaissances in the early years, but because they were discovered quickly, the losses were relatively small.
In contrast, it took several months for ltcm to discover that its model performed differently in reality than in backtests. By the time they first noticed a major model failure, it was too late; the company was on the verge of bankruptcy due to months of bad exposure that had accumulated.
Restrained growth and scale control
Compared with the huge success in performance, renaissance's product line has grown very little.
renaissance started out trading commodities and currencies, and by the mid-1980s it had mastered how to create sustainable but relatively small profits. To grow, it needs to tap into a bigger market: stocks. Renaissance has been testing stock trading algorithms for nearly a decade and found many that succeeded in backtesting but failed in the real world.
Some outstanding talents have left due to the frustration caused by constant attempts and failures. Some people think that Renaissance is too cautious about stocks, while others think that the stock problem cannot be solved. Fortunately, the core team of scientists persisted, and they finally found an algorithm that works in reality. This is the result of hundreds of incremental tweaks and bug fixes, not one moment of “aha.” Renaissance has slowly expanded this stock algorithm over the years.
Even as it scales up its equities business, Renaissance has taken a balancing act to achieve modest growth. In 2003, Simons concluded that Renaissance was managing more money than its model could comfortably allocate, so the company stopped taking money from outside investors, a decision that remained unchanged for years.
In contrast, ltcm's product types can be said to be in full bloom. It initially focused on sovereign debt convergence trading, and other firms took note of its success and began copying its trades. Within 24 months of its founding, the influx of competition drove down its profits, and ltcm looked elsewhere, subsequently investing heavily in companies with dual-class structures distinct from bonds and in merger arbitrage.
In fact, ltcm failed to demonstrate sustainable profitability in its core products and moved to new and even unproven products. This move is regarded as a sign of success by the outside world and the media.
For time series forecasting products, the best way to verify that success is real (and not just luck) is to demonstrate real-world accuracy over a long period of time. renaissance does this, ltcm does not.
Never doubt a model because of your intuition, and never completely trust a model.
ltcm was founded with a focus on one product: algorithmically generated sovereign debt convergence trades. The company believes in the power of its original algorithm—it’s based on the work of two Nobel Prize winners and has an enviable track record. Unfortunately, this record was short-lived. Based on very few results (due to the long-term nature of the trade), it is indistinguishable from a lucky streak. As a result, ltcm completely believed in its model.
has expanded from convergence transactions to merger arbitrage and dual-class companies, forcing ltcm to rely more on the intuition of its partners.
Of course, ltcm built models for these new products. However, there is no model to allocate risk between different products (asset classes), this part of the work is done by committees. Risk allocation ultimately reflects committee members' beliefs and internal politics.
For example, the head of ltcm's London office had influence within the company and was passionate about a particular Royal Dutch Shell deal. Therefore, despite the high concentration of risks and lack of successful experience in this transaction, ltcm invested more than 2 billion US dollars.
Let’s look at renaissance.
Simons was frustrated many times early in his investing career by ignoring algorithmic advice, and along the way he learned that you can and should recognize when a model has serious problems, stop trading on its recommendations, and fix them. At the same time, he also realized that if he was generally uncomfortable with the model's decisions, but the model's decisions were not obviously wrong, then his intuition would not produce better results.
In renaissance, it is commonplace to turn off a model, debug it, and then turn it back on again, but suggestions to overturn a model are irreversible; any reasons for overturning a decision must either be embedded in the model itself, or be discarded.
renaissance also benefits from its single trading model, which is responsible for all decisions for the company’s flagship fund. In software development, monolithic codebase has become a pejorative term, but it brings an interesting advantage to renaissance: it forces distribution trade-offs between products to be managed algorithmically. This differs from LTCM’s hybrid human-machine model, where the model picks the assets but a committee decides how to allocate risk among different asset classes.
Differentiated Data
The final factor that separates renaissance from ltcm is renaissance's focus on differentiated data.
renaissance started out by focusing on a relatively common data type: daily closing prices for commodities and currencies. This is data that other traders will also use. But renaissance realizes that differentiated data doesn’t just mean different types of data, it also means higher quality and broader coverage of the same data type compared to competitors.
Actually, it makes a lot of sense to focus on the same data types as your competitors.
The fact that all other trading firms value this data type shows that this data type provides a strong signal. Moreover, it is often easier and more cost-effective to amplify a known valuable signal than to find a completely new signal.
renaissance amplifies the signal by gathering daily price data further back than competitors, integrating data from more sources, and cleaning and cross-checking the data to eliminate erroneous values.
As the company grew, renaissance began to collect new data types: from daily closing prices to intraday prices, to company financial data, to free-text news reports...
In hindsight, it is easy to discount the advantages of renaissanceltcm Compare the shortcomings.But renaissance is not perfect, and ltcm also has some very innovative and valuable ideas that were quietly adopted by other companies after its collapse.
The real world is complex, but there is a simple rule of thumb that applies even to startups that have nothing to do with finance or algorithms: get the core product right before scaling up. Don’t confuse accolades from the media, investors, or even customers with the ability to consistently turn a profit. Ultimately, this is what ultimately allows a business to control its own destiny.
This article represents the author's views only.
-
-
-
-
-
-
-
-
-
-