SecProbe.io

CORRESP
1
filename1.htm

February 14, 2023

Ms. Alison White

U.S. Securities and Exchange Commission

100 F Street, NE

Washington, DC 20549

    Re:
    EA Series Trust (the “Trust”)

Post-Effective Amendment No. 202 to
the Registration Statement on Form N-1A (the

“Amendment”) File Nos.: 333-195493 and 811-22961

Euclidean Fundamental Value ETF

Dear Ms. White:

This correspondence responds to comments to the
Trust 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 Euclidean Fundamental Value ETF (the “Fund”),
a proposed new series of the Trust. For your convenience, the comments have been reproduced with responses following each comment. Capitalized
terms not otherwise defined have the same meaning as in the Amendment.

 1. Comment: In the Principal Investment Strategies section, please clarify how Step 4 of the
investment process identifies potential value traps. In addition, explain what, if any, data and/or analysis is used in Step 4 that is
not listed in Step 2 of the investment process.

Response: The
Fund confirms that the historical fundamental and market input data used by the model is the same input data for Step 2 and Step 4 of
the investment process. However, the Step 2 model was designed and is used to forecast future earnings (and uncertainty in those
earnings) whereas the model in Step 4 was designed and is used to forecast the probability of a company being a value-trap. See
the italicized text below for our changes in response to the above comment.

Step
4: Next, the model is used to identify and screen out potential “value-traps”. A potential value trap is defined
as a company that is considered inexpensive based on current valuation multiples but is likely to have bottom decile price performance
over the subsequent year. As in Step 2, the model takes as input historical company data, however, in Step 4 the model’s
output is the probability of a company not being a value-trap. This probability is then multiplied by the earnings yield alue from Step
3 and the Investment Universe is re-sorted such that, companies with a high probability of being a value-trap rank lower than companies
with a low probability of being a value-trap, all else being equal. This final ranking is then used as input for Step 5. The model
is implemented as a neural network and was developed using standard supervised machine learning techniques. The historical company data
used as input to the model in Step 4 includes financials such as income and cashflow statements, balance sheet data, short interest,
historical stock price change data, industry classification and market size category. The scope and types of input data used in Step
4 may change over time when it is found that additional types of data improve the accuracy of the model’s ability to predict
whether a company is or is not a value-trap.

11300 Tomahawk Creek Pkwy, Suite 310  ●
 Leawood, KS 66211

Practus, LLP   ●
  Practus.com

In addition, the Fund’s
sub-adviser, Euclidean Technologies Management, LLC (“Euclidean” or the “Sub-Adviser”), offers the following additional
background information about Euclidean, and the firm’s development of its investment models.

We (Euclidean) have
written an in-depth white paper that is publicly available on how the model in Step 4 identifies potential value traps. The paper can
be accessed here:

https://www.euclidean.com/value-traps-and-deep-learning

The Sub-Adviser
has developed the model to identify value-traps (defined as a company that is considered inexpensive based on current fundamentals but
is likely to have bottom decile price performance over the subsequent year) using standard supervised machine learning techniques. In
our approach, the model is implemented as an artificial neural network. The neural network represents an input-output mapping
(i.e., it maps input data to output data). Before learning begins, historical input and target output data is collected and the mapping
of input to output is unknown. The goal of supervised learning is to learn (via an algorithmic training process) the unknown mapping
such that when the neural network is presented with an particular set of inputs it produces the correct outputs (where correct means
the neural network output does not differ than the target output). In the process of training the neural network, the neural network
parameters are iteratively adjusted via the learning algorithm so that the difference between the model outputs and the target training
data outputs is minimized.

    2

In the value-trap model,
the input data is historical fundamental data about companies (e.g., data that can be found in a company’s income statement, balance
sheet, and cashflow statement), company historical stock price change data, company industry classification, company market capitalization
size category, and company total shares sold short divided by shares outstanding (short interest). The target output data is +1.0 (plus
one) or -1.0 (negative one), depending on whether the particular company, at a particular point-in-time, is a value-trap (+1.0) or not
a value-trap (-1.0). This setup is more easily visualized by the following table:

Each row represents
a company at a different point in time (where time is incremented in months) and the company’s input data and target output data
at that particular time. The data in the columns below “Fundamental Data at Time t” is the input data and the data
in the column below “Outcomes at Time t+12” is the target output data (value-trap or not value-trap). The goal of the
learning algorithm is teach the neural network to produce the data in the output column when presented with the data in the input columns.

The key difference between
the value-trap model in Step 4 and the earnings forecast model in Step 2 is that the two models are attempting to forecast different things.
In the case of the value-trap model, the model is taught to forecast whether a company, at a particular point in time, is likely to be
a bottom decile stock market performer over the subsequent year (i.e., a value-trap). However, the target output of the model in Step
2 is the company’s future earnings in twelve months time. So, in the case of the value-trap model, it is attempting to forecast
something about a company’s future stock price (how well it performs relative to other stocks). Whereas, in the case of the earnings
forecast model, it is attempting to forecasts something about a company’s future fundamentals, and in particular its earnings.

 2. Comment: In the Principal Investment Strategies section, the Staff notes that the Fund mentions
in Step 5 of the investment process that the Sub-Adviser performs a manual evaluation of the data used in the investment model’s
assessment process. It further notes that if the Sub-Adviser believes the data to be inaccurate or new data has become available that
may impact the investments risks for the Fund the Sub-Adviser may override the model. With a view to enhance disclosure under Item 9 of
Form N-1A, further explain how the model was developed and validated. Please address the data and relationships your machine learning
model relies on. Explain how the Sub-Adviser ensures such relationship(s) hold over time.

    3

Response: The
Fund has added the following information to the Item 9 discussion to address the noted items.

Additional Information
About the Sub-Adviser’s Models. The models are implemented as an artificial neural network, and each model is trained on data
through a supervised machine learning process. A supervised machine learning process is an automated process, implemented as a computer
program, where an input-output mapping is learned from data. The two models are different as they serve different functions and hence
implement different input-output mappings.

The first model is designed
as an earnings forecasting and uncertainty quantification model (See Step 2 of the investment process). Using standard deep machine learning
techniques, the model was trained to map historical data about every company in the Investment Universe, at each point-in-time, to the
company’s one-year ahead future earnings. It was simultaneously trained to quantify the uncertainty (using the maximum likelihood
technique) in the forecast future earnings. This training process was executed in an iterative manner starting on data in the year 1970
and going through 1999. The model was validated on out-of-sample data beginning in the year 2000 and going through 2019.

The second model predicts
whether an investment recommended by the first model is likely to be a value-trap (See Step 4 of the investment process). Again, using
standard deep machine learning techniques, the model was trained to map historical data about every company in the Investment Universe,
at each point-in-time, to the probability that the company is a value-trap (defined as having a bottom decile stock price return over
the subsequent year). This training process was executed in an iterative manner starting on data in the year 1970 and going through 1999.
The model was validated on out-of-sample data beginning in the year 2000 and going through 2019.

The relationships the
machine learning models rely on are between the input and output data as learned by the models. Both models use the same input data: historical
fundamental data about companies (e.g., data that can be found in a company’s income statement, balance sheet, and cashflow statement),
company historical stock price change data, company industry classification, company market capitalization size category, and company
total shares sold short divided by shares outstanding (short interest). The output data for the earnings’ forecast model is a company’s
future earnings and the output data for the value-trap model is whether a company is or is-not a value-trap (defined as having a bottom
decile stock price return over the subsequent year). Therefore, each model has learned a separate input-output relationship.

These input-output relationships
can change over time (called a non-stationary data distribution) and the Sub-Adviser’s processes take this into account. This is
accomplished by periodically updating the models so that each model is trained through the present, incorporating shifts in the input-output
relationships that may have occurred since the model was last trained. However, the Sub-Adviser has observed in practice by researching
how these relationships have evolved over the last 60 years, that changes happen at a very slow pace and likely only necessitating model
updates every three to five years. The overall performance of each model was validated with an investment simulator over the out-of-sample
period 2000 through 2019.

    4

The investment process,
as described in Step 2 through Step 4 in the Principal Investment Strategies section, automatically generate a list of investments for
the Fund. The manual oversight by the Sub-Adviser, as described in Step 5 of the Principal Investment Strategies section, should be viewed
as a final check of the models, and does not allow for the Sub-Adviser to choose an investment for the Fund that is not considered a good
investment by the models. There are two broad instances when Step 5 of the investment process plays a role: (1) the model(s) is/are receiving
erroneous data to generate a specific recommendation (i.e., “garbage in, garbage out”) and (2) there exists highly relevant
negative information about the company that the model(s) does not have access to. To further explain the first instance, each model relies
on data provided by third party data vendors (primary Standard & Poor’s Compustat and CapitalIQ data services). The data includes
information compiled from company SEC filings, exchanges, and other sources. In the case of SEC filings, the data is transcribed from
the filing into a highly structured and normalized data format by the data provider, often very shortly after the filing has occurred.
There can be errors, caused either by human or computer programs, in this transcription process. The company can also, shortly after a
filing, say that the filing needs to be restated. The same can be true of market data that is provided by an exchange. In these cases,
when the Sub-Adviser knows a model is making a recommendation on erroneous input data, the Sub-Adviser cannot be confident that the output
of the model is of any value and therefore puts a hold on considering the particular company in question as an investment until the data
error has been resolved. The second instance where Sub-Adviser oversight plays a role as part of Step 5 of the investment process is when
there is highly relevant negative information about the company that the models do not have access to. Examples include, but are not limited
to, the CEO being indicted of a financial crime, the company has announced undisclosed financial misrepresentations, the company files
for bankruptcy, or the company is losing a customer that makes up a substantial portion of its revenue (e.g., 30%). This manual process
is not designed or used by the Sub-Adviser to routinely override the models, but simply provides an additional layer of protection to
the integrity of the models.

 3. Comment: In the Principal Investment Strategies section, please clarify what you mean by
“machine learning” in Step 2. The investment process appears to be based on fundamental stock analysis so how does machine
learning fit into the fundamental investment process. Explain how the Sub-Adviser’s investment model is different from others used
in the industry.

Response: The
Fund has modified the discussion in Step 2 to further explain what it means by machine learning. See the revised Step 2 discussion below.
In addition, as part of the Fund’s response to this comment the Fund refers you to Response 2 (above) and the additional information
included in the Item 9 discussion regarding the Sub-Adviser’s models.

Step
2: In this Step, the Sub-Adviser applies its machine learning based model to generate a discounted forecast of next year’s earnings
for each company within the Investment Universe. This discounted earnings forecast is then used as input for Step 3. The model
is implemented as a neural network and was developed using standard supervised machine learning techniques. The model was trained to implement
an input-output mapping which maps historical company data to future (forecast) company earnings. The model was also trained to estimate
the uncertainty in its forecasts and this uncertainty is used to discount the earnings forecast such that the final product of the model
in Step 2 is a discounted earnings forecast. The historical company data used as input to the model in Step 2 includes
financials such as income and cashflow statements, balance sheet data, short interest, historical stock price change data, industry classification
and market size category. The scope and types of input data used in Step 2 may change over time when it is found that additional
types of data improve the accuracy of the discounted earnings forecast generated in Step 2.

    5

As it relates to the
comment regarding how machine learning fits into what appears to be a fundamental stock analysis strategy, the Fund first refers you to
Response 2 above where additional information about the Sub-Adviser’s Model is presented. The Fund and the Sub-Adviser are not clear
on what part of Step 2 of the investment process implies that the “investment process appears to be based on fundamental stock analysis.”
Step 2 uses the machine learning based model described herein to generate a discounted earnings forecast for each company in the investment
universe. That discounted earnings forecast is then used to calculate an earnings yield factor by dividing the discounted earnings forecast
by the comp