Correspondence 0001592900-24-000929 from EA Series Trust (CIK 0001592900)
EA Series Trust (CIK 0001592900)
Date: June 4, 2024 · CIK: 0001592900 · Accession: 0001592900-24-000929
AI Filing Summary & Sentiment
File numbers found in text: 333-195493, 811-22961
Letter
Washington, DC 20549 Post-Effective Amendment No. 325 to the Registration Statement on Form N-1A (the “Amendment”) File Nos.: 333-195493 and 811-22961 Draco Evolution AI ETF
RE: EA Series Trust (the “Trust”)
Dear Ms. Rossotto:
The Trust is filing correspondence in order to supplement several of the Trust’s original responses, filed on May 6, 2024, to the comments received by the undersigned from the staff of the U.S. Securities and Exchange Commission (the “Staff” of the “Commission”) with respect to the Amendment relating to the Draco Evolution AI ETF (the “Fund”), a proposed new series of the Trust. For your convenience, the Trust has included the Staff’s original comment along with the Trust’s supplemented response. Unless listed below, the Trust’s original responses to the Staff’s comments as filed on May 6, 2024 remain unchanged. The Trust has included as Exhibit A the pages from the prospectus that have been updated to address the Staff’s comments.
1.Comment: (Original Comment 4 – May 6, 2024 Correspondence) Provide additional details in correspondences on how the proprietary AI model was developed for the investment process including:
•A discussion of signal to noise ratio and interpretability factors;
•The types of data inputs entered;
•The types of signals or output received;
•How is the output unbiased;
•How is the data validated for accuracy and reliability and who performs the validation;
•What specific outputs were considered;
•Any limitations on the use of the AI technology and the extent to which a portfolio manager will override the AI signals
Provide additional disclosure in the prospectus regarding the following:
•The use of incomplete data or erroneous data
•Potential for changes in relationships for data sets that could result in less useful outcomes or recommendations and/or potential for loss due to coding or logic errors and malware or other potential technology degradation associated with the use of technology.
Response: The Sub-Adviser to the Fund provides the following information in response to the comments regarding its AI model.
•The signal-to-noise ratio (SNR) quantifies the quality of information in a system by comparing the strength of the useful signal (relevant information) to the level of background noise (irrelevant information). The higher the SNR the better the AI model is in distinguishing relevant information from irrelevant information, resulting in more robust predictions. In analyzing the level of noise present in the strategy, the Sub-Adviser measures, such things as, the volatility of returns, calculating standard deviations or it may use other statistical techniques to quantify the randomness or unpredictability of the Fund’s investment strategy. For example, calculating the standard deviation of returns over a certain period provides an estimate of the volatility or noise level in the strategy. The Sub-Adviser then calculates the Sharpe ratio (return/volatility>1) to assess the quality of the predictions.
11300 Tomahawk Creek Pkwy, Suite 310 ● Leawood, KS 66211
Practus, LLP ● Practus.com
•To strengthen the SNR, the Sub-Adviser will value and focus on Interpretability, which refers to the degree to which a trading strategy’s decisions and actions can be understood and explained by humans, especially in economic logic, by building the AI model based on methods like Modern Portfolio Theory and Efficient Frontier, which is a mathematical framework for assembling a portfolio of assets such that the expected return is maximized for a given level of risk. It is a formalization and extension of diversification in investing, the idea that owning different kinds of financial assets is less risky than owning only one type. Its key insight is that an asset's risk and return should not be assessed by itself, but by how it contributes to a portfolio's overall risk and return.
•The AI model’s inputs are based on high frequency data such as stock prices, trading volumes, and proprietary indicators while the inputs for the quantitative model encompasses lower frequency, macroeconomic data, such as average weekly hours worked, applications for unemployment insurance, new orders for consumer goods and materials, new orders for capital goods, new building permits, the spread between long and short interest rates, the inflation-adjusted money supply, and average consumer expectations for business conditions.
•The AI model utilizes XGboost, an open-source gradient boosting model, to predict whether the next 6 months will be good or bad for markets (bullish or bearish). The AI model is trained using two years of daily pricing and volume data from 14 different ETFs. These ETFs cover a variety of asset classes including but not limited to US Equities, US Corporate Bonds, US Treasury Bonds, and Gold. The pricing series from these funds are processed into a variety of technical indicators that are used as features to train the AI model. Based on this 2-year window of high-frequency data, the AI model determines whether the environment is currently bullish or bearish (AI output).
•Output from the AI model is augmented by a quantitative model. This model relies on monthly macroeconomic data going back to 2007. The macro model attempts to decipher the long-term economic environment and further refines weights set by the AI model.
•The Sub-Adviser performs an out-of-the-sample test to ensure that the output is unbiased. In machine learning, an “out-of-sample test” refers to evaluating the performance of a trained model on data that it has not seen during the training process. During the training phase, a machine learning model learns patterns and relationships in a dataset. However, the true test of its effectiveness is its ability to generalize to new, unseen data so the out-of-sample test involves feeding new data, which the model hasn't encountered before, to the trained model and assessing its performance. This test helps determine how well the model can make accurate predictions or classifications on data it wasn't trained on, providing insights into its real-world applicability and generalization capabilities.
•The Sub-Adviser validates the data every time they use it by comparing data from different sources. By using multiple sources for the data, the Sub-Adviser believes this reduces the likelihood of inaccurate or unreliable data. The Sub-Adviser’s data resources are from TD Ameritrade and Yahoo finance and to a lesser extent from Interactive Brokers. If the Sub-Adviser encounters data inconsistencies, it uses the following experience-based criteria to determine which data source to trust - whether there is any abnormal slippage in stock price information. For example, if the Sub-Adviser observes that the price of a stock at time “t” in minute “k’’ is 200, but at t+1 it abnormally drops below 100, and then returns to around 200 at t+2, this is a clear indication of an anomaly in the data provider’s database. The Sub-Adviser would not use data showing such abnormal price movements. If the data differences are very minor and do not affect the operation of our AI model, they are negligible and can be ignored.
•The use of AI technology to assist in managing a fund is like any other fund where humans and technology are involved in the process. For example, humans either directly or through the use of technology can (among other things) make errors, misinterpret market data and/or allocate investments unsuccessfully. The use of AI technology in this process does not eliminate the possibility for errors or unsuccessful investment results. The Fund has included the following risk disclosures to inform shareholders that AI based strategies are no different from traditional strategies in that risk is involved with an investment in the Fund: AI Model
Risk, Quantitative Security Selection Risk and Management Risk. The Fund believes these three risks coupled with the other Principal Risk provide shareholders with a robust risk discussion.
•The Sub-Adviser’s AI model monitors for risk based on the underlying market activity of its holdings. As currently disclosed in the prospectus, the Sub-Adviser’s AI model monitors the Fund’s portfolio throughout the day and, if the Fund’s intraday drawdown exceeds 10% or when the Fund’s drawdown exceeds 20% from its all-time high the AI model will be paused, and the Sub-Adviser will be alerted. The Sub-Adviser will assess the situation and decide what steps, if any, to take as it relates to the portfolio.
•The Fund has expanded its risk disclosure to address the use of incomplete data or erroneous data and the potential for changes in relationships for data sets that could result in less useful outcomes or recommendations and/or potential for loss due to coding or logic errors and malware or other potential technology degradation associated with the use of technology.
2. Comment: (Original Comment 11 – May 6, 2024 Correspondence) In correspondence, explain what makes the investment process employed by the Fund AI and not quantitative.
Response: The Sub-Adviser’s model (collectively called the “Draco Model”) consists of two parts: the AI model and the macroeconomic quantitative model. The AI model is trained with two years of high frequency pricing data and makes a short-term prediction on market regime (bearish or bullish). This step is responsible for determining about 70% of the strategy’s weight. While the quantitative model uses lower frequency macroeconomic data to predict the long-term market outlook. Weights are adjusted based on this prediction. Finally, the strategy’s weights are adjusted based on current market volatility.
The AI model determines which asset classes the strategy has exposure to, depending on its market condition predictions. About 70% of its weight is allocated in the AI model step.
3. Comment: (Original Comment 15 – May 6, 2024 Correspondence) In the Principal Investment Strategies section, the Fund states that the AI driven algorithms “automatically invest, divest and rebalance the Fund’s portfolio”. Explain how the Fund’s automatic trading system monitors industry concentration.
Response: The Sub-Adviser’s trading system (the Draco Model) consists of two parts: a) AI involves using an AI model to analyze and predict the optimal portfolio for the Fund based on its prediction of short-term market conditions. b) Quantitative model involves a review of several macroeconomic indicators (e.g., GDP, CPI, PCE, etc.) and predicts the long-term economic environment and further refines weights set by the AI model. Once the Draco Model generates a suggestion of asset allocation weights (output), the system automatically checks to ensure the portfolio meets Fund’s investment strategy and risk requirements (e.g. investment caps in place for such requirements as industry concentration). Next, the portfolio manager performs a manual check on the portfolio. The trading system is designed to rebalance the Fund’s portfolio in response to its market predictions.
4.Comment: (Original Comment 23 – May 6, 2024 Correspondence) Please disclose in correspondence whether the Sub-Adviser has used the AI model to manage money for any other clients/accounts over a full market cycle. If not, consider expanding the risk disclosure to note the newness of the model.
Response: The Sub-Adviser has been using the Draco Model to manage money since 2020 for hedge funds and SMAs. The Fund believes this period of time removes the need to disclose the newness of the model.
If you have any questions regarding the above responses, please do not hesitate to contact me at (513) 304-5605 or Wade.Bridge@practus.com.
Sincerely,
/s/ Wade Bridge
Show Raw Text
CORRESP
1
filename1.htm
Document
June 4, 2024
Ms. Karen Rossotto
U.S. Securities and Exchange Commission
100 F Street, NE
Washington, DC 20549
RE: EA Series Trust (the “Trust”)
Post-Effective Amendment No. 325 to the Registration Statement on Form N-1A (the “Amendment”) File Nos.: 333-195493 and 811-22961
Draco Evolution AI ETF
Dear Ms. Rossotto:
The Trust is filing correspondence in order to supplement several of the Trust’s original responses, filed on May 6, 2024, to the comments received by the undersigned from the staff of the U.S. Securities and Exchange Commission (the “Staff” of the “Commission”) with respect to the Amendment relating to the Draco Evolution AI ETF (the “Fund”), a proposed new series of the Trust. For your convenience, the Trust has included the Staff’s original comment along with the Trust’s supplemented response. Unless listed below, the Trust’s original responses to the Staff’s comments as filed on May 6, 2024 remain unchanged. The Trust has included as Exhibit A the pages from the prospectus that have been updated to address the Staff’s comments.
1.Comment: (Original Comment 4 – May 6, 2024 Correspondence) Provide additional details in correspondences on how the proprietary AI model was developed for the investment process including:
•A discussion of signal to noise ratio and interpretability factors;
•The types of data inputs entered;
•The types of signals or output received;
•How is the output unbiased;
•How is the data validated for accuracy and reliability and who performs the validation;
•What specific outputs were considered;
•Any limitations on the use of the AI technology and the extent to which a portfolio manager will override the AI signals
Provide additional disclosure in the prospectus regarding the following:
•The use of incomplete data or erroneous data
•Potential for changes in relationships for data sets that could result in less useful outcomes or recommendations and/or potential for loss due to coding or logic errors and malware or other potential technology degradation associated with the use of technology.
Response: The Sub-Adviser to the Fund provides the following information in response to the comments regarding its AI model.
•The signal-to-noise ratio (SNR) quantifies the quality of information in a system by comparing the strength of the useful signal (relevant information) to the level of background noise (irrelevant information). The higher the SNR the better the AI model is in distinguishing relevant information from irrelevant information, resulting in more robust predictions. In analyzing the level of noise present in the strategy, the Sub-Adviser measures, such things as, the volatility of returns, calculating standard deviations or it may use other statistical techniques to quantify the randomness or unpredictability of the Fund’s investment strategy. For example, calculating the standard deviation of returns over a certain period provides an estimate of the volatility or noise level in the strategy. The Sub-Adviser then calculates the Sharpe ratio (return/volatility>1) to assess the quality of the predictions.
11300 Tomahawk Creek Pkwy, Suite 310 ● Leawood, KS 66211
Practus, LLP ● Practus.com
•To strengthen the SNR, the Sub-Adviser will value and focus on Interpretability, which refers to the degree to which a trading strategy’s decisions and actions can be understood and explained by humans, especially in economic logic, by building the AI model based on methods like Modern Portfolio Theory and Efficient Frontier, which is a mathematical framework for assembling a portfolio of assets such that the expected return is maximized for a given level of risk. It is a formalization and extension of diversification in investing, the idea that owning different kinds of financial assets is less risky than owning only one type. Its key insight is that an asset's risk and return should not be assessed by itself, but by how it contributes to a portfolio's overall risk and return.
•The AI model’s inputs are based on high frequency data such as stock prices, trading volumes, and proprietary indicators while the inputs for the quantitative model encompasses lower frequency, macroeconomic data, such as average weekly hours worked, applications for unemployment insurance, new orders for consumer goods and materials, new orders for capital goods, new building permits, the spread between long and short interest rates, the inflation-adjusted money supply, and average consumer expectations for business conditions.
•The AI model utilizes XGboost, an open-source gradient boosting model, to predict whether the next 6 months will be good or bad for markets (bullish or bearish). The AI model is trained using two years of daily pricing and volume data from 14 different ETFs. These ETFs cover a variety of asset classes including but not limited to US Equities, US Corporate Bonds, US Treasury Bonds, and Gold. The pricing series from these funds are processed into a variety of technical indicators that are used as features to train the AI model. Based on this 2-year window of high-frequency data, the AI model determines whether the environment is currently bullish or bearish (AI output).
•Output from the AI model is augmented by a quantitative model. This model relies on monthly macroeconomic data going back to 2007. The macro model attempts to decipher the long-term economic environment and further refines weights set by the AI model.
•The Sub-Adviser performs an out-of-the-sample test to ensure that the output is unbiased. In machine learning, an “out-of-sample test” refers to evaluating the performance of a trained model on data that it has not seen during the training process. During the training phase, a machine learning model learns patterns and relationships in a dataset. However, the true test of its effectiveness is its ability to generalize to new, unseen data so the out-of-sample test involves feeding new data, which the model hasn't encountered before, to the trained model and assessing its performance. This test helps determine how well the model can make accurate predictions or classifications on data it wasn't trained on, providing insights into its real-world applicability and generalization capabilities.
•The Sub-Adviser validates the data every time they use it by comparing data from different sources. By using multiple sources for the data, the Sub-Adviser believes this reduces the likelihood of inaccurate or unreliable data. The Sub-Adviser’s data resources are from TD Ameritrade and Yahoo finance and to a lesser extent from Interactive Brokers. If the Sub-Adviser encounters data inconsistencies, it uses the following experience-based criteria to determine which data source to trust - whether there is any abnormal slippage in stock price information. For example, if the Sub-Adviser observes that the price of a stock at time “t” in minute “k’’ is 200, but at t+1 it abnormally drops below 100, and then returns to around 200 at t+2, this is a clear indication of an anomaly in the data provider’s database. The Sub-Adviser would not use data showing such abnormal price movements. If the data differences are very minor and do not affect the operation of our AI model, they are negligible and can be ignored.
•The use of AI technology to assist in managing a fund is like any other fund where humans and technology are involved in the process. For example, humans either directly or through the use of technology can (among other things) make errors, misinterpret market data and/or allocate investments unsuccessfully. The use of AI technology in this process does not eliminate the possibility for errors or unsuccessful investment results. The Fund has included the following risk disclosures to inform shareholders that AI based strategies are no different from traditional strategies in that risk is involved with an investment in the Fund: AI Model
2
Risk, Quantitative Security Selection Risk and Management Risk. The Fund believes these three risks coupled with the other Principal Risk provide shareholders with a robust risk discussion.
•The Sub-Adviser’s AI model monitors for risk based on the underlying market activity of its holdings. As currently disclosed in the prospectus, the Sub-Adviser’s AI model monitors the Fund’s portfolio throughout the day and, if the Fund’s intraday drawdown exceeds 10% or when the Fund’s drawdown exceeds 20% from its all-time high the AI model will be paused, and the Sub-Adviser will be alerted. The Sub-Adviser will assess the situation and decide what steps, if any, to take as it relates to the portfolio.
•The Fund has expanded its risk disclosure to address the use of incomplete data or erroneous data and the potential for changes in relationships for data sets that could result in less useful outcomes or recommendations and/or potential for loss due to coding or logic errors and malware or other potential technology degradation associated with the use of technology.
2. Comment: (Original Comment 11 – May 6, 2024 Correspondence) In correspondence, explain what makes the investment process employed by the Fund AI and not quantitative.
Response: The Sub-Adviser’s model (collectively called the “Draco Model”) consists of two parts: the AI model and the macroeconomic quantitative model. The AI model is trained with two years of high frequency pricing data and makes a short-term prediction on market regime (bearish or bullish). This step is responsible for determining about 70% of the strategy’s weight. While the quantitative model uses lower frequency macroeconomic data to predict the long-term market outlook. Weights are adjusted based on this prediction. Finally, the strategy’s weights are adjusted based on current market volatility.
The AI model determines which asset classes the strategy has exposure to, depending on its market condition predictions. About 70% of its weight is allocated in the AI model step.
3. Comment: (Original Comment 15 – May 6, 2024 Correspondence) In the Principal Investment Strategies section, the Fund states that the AI driven algorithms “automatically invest, divest and rebalance the Fund’s portfolio”. Explain how the Fund’s automatic trading system monitors industry concentration.
Response: The Sub-Adviser’s trading system (the Draco Model) consists of two parts: a) AI involves using an AI model to analyze and predict the optimal portfolio for the Fund based on its prediction of short-term market conditions. b) Quantitative model involves a review of several macroeconomic indicators (e.g., GDP, CPI, PCE, etc.) and predicts the long-term economic environment and further refines weights set by the AI model. Once the Draco Model generates a suggestion of asset allocation weights (output), the system automatically checks to ensure the portfolio meets Fund’s investment strategy and risk requirements (e.g. investment caps in place for such requirements as industry concentration). Next, the portfolio manager performs a manual check on the portfolio. The trading system is designed to rebalance the Fund’s portfolio in response to its market predictions.
4.Comment: (Original Comment 23 – May 6, 2024 Correspondence) Please disclose in correspondence whether the Sub-Adviser has used the AI model to manage money for any other clients/accounts over a full market cycle. If not, consider expanding the risk disclosure to note the newness of the model.
Response: The Sub-Adviser has been using the Draco Model to manage money since 2020 for hedge funds and SMAs. The Fund believes this period of time removes the need to disclose the newness of the model.
If you have any questions regarding the above responses, please do not hesitate to contact me at (513) 304-5605 or Wade.Bridge@practus.com.
Sincerely,
/s/ Wade Bridge
Wade Bridge
Counsel
3