Finance Research Letters 58 (2023) 104580

Contents lists available at ScienceDirect

Finance Research Letters

journal homepage: www.elsevier.com/locate/frl

What if ChatGPT were a quant asset manager

Jang Ho Kim a, b, *
a
Department of Industrial and Management Systems Engineering, College of Engineering, Kyung Hee University, 1732, Deogyeong-daero, Giheung-
gu, Yongin-si, Gyeonggi-do 17104, South Korea
b
Department of Big Data Analytics, Graduate School, Kyung Hee University, 1732, Deogyeong-daero, Giheung-gu, Yongin-si, Gyeonggi-do 17104,
South Korea

A R T I C L E I N F O A B S T R A C T

Keywords: Even though large language models such as ChatGPT are not specifically trained for analyzing
ChatGPT asset returns or recommending stocks, it may still provide additional insight into making in­
Portfolio optimization vestment decisions. In this study, we propose a quantitative investment approach that in­
Economic indicators
corporates recommendations from ChatGPT. Based on ChatGPT’s general understanding of
Asset allocation
economy and financial market movements, we ask ChatGPT to recommend asset classes under
various economic conditions. Our empirical results show that asset class recommendations based
on economic indicators of ChatGPT can improve portfolio efficiency.

1. Introduction

While general large language models such as OpenAI’s ChatGPT are opening potential to various research domains and applica­
tions, it is yet to be trained specifically to become an expert in analyzing financial markets or recommending stocks. Nonetheless, it
portrays a general understanding of economy and finance at the macro level. In this study, we show that even this basic level of
knowledge can add value to making investment decisions. In particular, we perform empirical analysis where ChatGPT is given the task
of a quant asset manager that recommends asset classes to invest under various economic conditions.
Several studies already use ChatGPT for research in the finance domain. For example, Dowling and Lucey (2023) demonstrate how
ChatGPT can assist with finance research, and Lo et al. (2023) discuss how natural language processing has been affecting the finance
industry, with ChatGPT as a coauthor. Studies that focus more on market analysis include Aldridge (2023), which illustrates the
evolution from linear regression to ChatGPT and beyond for analyzing stock returns, and Saggu and Ante (2023), which analyzes the
effects of launching ChatGPT to crypto assets. Recently, Romanko et al. (2023) and Ko and Lee (2023) explore ChatGPT for asset
selection and portfolio construction. While the two studies are related to ours since ChatGPT is utilized for making investment de­
cisions, we explore ChatGPT’s broad understanding of the relation between the economy and financial assets without testing its stock
selection and portfolio construction ability. Our key contribution is to demonstrate the use of ChatGPT as a component of quantitative
investment decision-making and show its effectiveness by proposing a quant portfolio management process that involves ChatGPT.
This process contains key elements of quant investing such as factor models and portfolio optimization but replaces asset managers’
involvement with ChatGPT. Our findings show the potential value and efficiency of incorporating recommendations on asset class by
ChatGPT.

* Correspondence to: Department of Industrial and Management Systems Engineering, College of Engineering, Kyung Hee University, 1732,
Deogyeong-daero, Giheung-gu, Yongin-si, Gyeonggi-do 17104, South Korea.
E-mail address: [email protected].

https://doi.org/10.1016/j.frl.2023.104580
Received 7 August 2023; Received in revised form 9 October 2023; Accepted 9 October 2023
Available online 11 October 2023
1544-6123/© 2023 Elsevier Inc. All rights reserved.
J.H. Kim Finance Research Letters 58 (2023) 104580

2. Methodology

Since the main focus of the study is to use GPT as a quant asset manager, we consider a process for constructing quantitative
portfolios that is based on risk factors, economic scenarios, and portfolio optimization1. Following many quant strategies, optimal
allocation weights are computed from portfolio optimization models. However, ChatGPT acts as a quant asset manager that determines
a list of asset classes to include in the portfolio, prior to performing portfolio optimization. Thus, ChatGPT only affects the list of asset
classes included in a portfolio but does not directly compute (or recommend) the weight allocated to each asset class. More specifically,
we ask ChatGPT to recommended asset classes given current economic conditions because ChatGPT will be able to learn macro relation
between economy and financial market movements. Making investment decisions per economic condition is an approach widely used
in scenario analysis of portfolios, where economic scenarios are defined with economic variables such as economic growth, unem­
ployment rate, and inflation (Czasonis et al., 2020; Kritzman et al., 2021; Gavell et al., 2022).
The overall process is summarized in panel (a) of Fig. 1. In step 0, significant factors (economic indicators) that explain market
movements are identified. Then, in step 1, the significant indicators are used to characterize the current economic condition (e.g.,
rising/falling inflation or strengthening/weakening US dollar). But since economic indicators cannot be directly invested, the list of
assets to invest under current economic condition is determined in step 2, and step 3 computes the optimal asset allocation from
portfolio optimization. Finally, in step 4, the portfolio is rebalanced to the optimal allocation until the next rebalancing date.
In our proposed portfolio process, ChatGPT is used in steps 0 and 2 where it chooses the significant economic indicators and decides
which assets to invest under various conditions of the selected indicators. For example, we request ChatGPT to recommend asset
classes when inflation is rising, interest rates are rising, and US dollar is strengthening. We stress that we are not assuming ChatGPT to
be a trained professional and not asking ChatGPT to predict asset returns. Since GPT is a language model, we are using the knowledge
of GPT that connects economic conditions to major asset classes; ChatGPT will be able to learn the macro relation between economic
conditions (or cycles) and asset class movements. Moreover, note that we do not ask ChatGPT to evaluate current market conditions
because this will require up-to-date information; ChatGPT only maps asset classes to economic conditions. Finally, while we explicitly
assign the task of an asset manager to ChatGPT, we do not affect the results of ChatGPT through conversating because the response of
ChatGPT will inevitably reflect the user’s inputs to a certain degree.
We also note that the process in Fig. 1 is closely related to quantitative models such as the All Weather strategy (Podolsky et al.,
2012). In the All Weather portfolio, which builds the foundation of risk parity models, the range of economic environments that in­
vestors may face are identified (e.g., rising economic growth and rising inflation), and risk is equally allocated to each economic
environment. Asset classes are mapped to economic environments so that equal allocation of risk to each environment can be achieved
through investing in the asset classes corresponding to each environment. Similarly, we use significant economic indicators for forming
economic environments. However, the key significance in our approach is to follow the recommendations of ChatGPT for identifying
economic indicators and mapping asset classes to economic conditions. Panel (b) of Fig. 1 describes how we implemented the proposed
portfolio process, which we further describe in the following section.
When computing optimal allocations, we test with several portfolio models including equally-weighted (Eqw), global minimum-
variance (GMV), mean-variance (M-V), and risk parity (RP) portfolios2. Eqw, GMV, and RP portfolios are more appropriate for
evaluating the significance of ChatGPT’s recommendation because they form more diversified allocations among candidate assets. M-V
portfolios with 5 % annualized mean return are also included to observe results when more aggressive (return-seeking) portfolios are
formed.

3. Data

Two types of data are retrieved from different sources in our experiment: economic data from Federal Reserve Economic Data
(FRED) and ETF returns from Yahoo Finance. Economic data are used for modeling economic conditions and trends in our backtesting.
We retrieved key economic indicators that are published with monthly frequency because it provides enough data points for analyzing
significance. Seven indicators are considered: unemployment rate, yield spread (10-year minus 2-year treasury rates), federal funds
rate, inflation rate, real GDP change rate, crude oil price, and dollar index. Monthly data from 2012 to July 2023 are retrieved, which
includes the most recent data at the time of the analysis.
The second type of data, ETF returns, are used for computing portfolio performance. Since we do not assume ChatGPT can
recommend individual stocks, we ask ChatGPT to recommend asset classes, and portfolios are invested at the asset class level. Thus,
portfolios invest in ETFs and daily returns are retrieved from 2013 to July 2023.

4. Empirical results: asset class selection from ChatGPT

In this section, we begin with our findings on asset class selection and present backtest results in Section 5. The actual interaction
with ChatGPT is included in the online appendix.
First, we ask ChatGPT to select significant economic indicators. In particular, when asked to identify three among seven indicators

1
See, for example, Greenberg et al. (2016) for the use of factors to asset allocation, Czasonis et al. (2020) for the use of economic scenarios for
analyzing portfolios, and Kim et al. (2021) for the use of optimization for making asset allocation decisions.
2
See Kim et al. (2021) and Qian (2011) for details on mean-variance optimization and risk parity, respectively.

2
J.H. Kim Finance Research Letters 58 (2023) 104580

Fig. 1. Quantitative investment process.

that are updated on a monthly basis (as shown below), interest rate (RATE), inflation (INF) and US dollar (USD) were chosen.
“Among the following, pick the three most significant indicators for modeling stock returns: {unemployment rate, yield spread, interest
rate, inflation, real GDP, oil price, US dollar}”
All combinations of rising and falling trends for the three indicators result in eight economic conditions. Therefore, we ask ChatGPT
to recommend asset classes that are good investments under each of the eight economic environments. For example, we ask ChatGPT to
recommend at most three asset classes by entering the following request3,
“Suppose you are an asset manager. Pick at most three asset classes that are good investments under the following economic conditions:
rising interest rate, rising inflation, and strengthening dollar.”
Recommendations by ChatGPT are summarized in Table 1 and further detailed in Table 24. As shown in Table 2, the responses by
ChatGPT included a total of 11 asset classes and we have matched asset classes with representative ETFs. In our proposed quant in­
vestment approach, the economic condition of the current month is determined (among the possible eight conditions) and the cor­
responding three asset classes shown in Table 1 are used for forming optimal portfolios. Allocation weights are optimized from
portfolio optimization models such as M-V or RP. Therefore, the list of three asset classes may change each month, but the investment is
always allocated among three ETFs.
Furthermore, in order to check the value of asset class selection of ChatGPT, we compare performance of portfolios that do not
incorporate ChatGPT. We form portfolios that invest in all 11 asset classes (consider all ETFs every month), and also form portfolios
that invest in five major asset classes, which are specified in Table 2. For both cases (11 or five asset class portfolios), we test with
various portfolio optimization models to provide a comprehensive comparison. The actual backtest results are presented next.

5. Empirical results: backtest

5.1. Backtest results in-sample

The in-sample backtest is performed from January 2013 to September 2021 because GPT is trained until this date. This in-sample
setting uses information up to the current month for making investment decisions at the beginning of the month. For example, the
economic condition for January 2013 is used for selecting three asset classes for January 2013 and the portfolio is assumed to invest in
these assets during January 2013. Furthermore, we use 1-, 3-, 6- or 12-month momentums for defining the economic condition (rising
or falling) for each month. For example, the 3-month momentums are shown in Table 3. Portfolio constructed from 1-, 3-, 6- and 12-
month economic trends are denoted as GPT1m, GPT3m, GPT6m, and GPT12m, respectively.
In order to observe the sole effects of asset class selection, we form Eqw portfolios that will not be affected by the choice of
optimization model. As shown in Fig. 2 and Table 4, the GPT portfolios except GPT12m show higher return, lower volatility, higher
Sharpe ratio, smaller maximum drawdown (MDD), smaller conditional value-at-risk (CVaR) compared to the Eqw portfolio of five
major asset classes. This clearly shows that the asset class recommendation of ChatGPT is valuable for forming more efficient in­
vestments. Note that the Eqw portfolio of 11 assets also incorporates recommendations of ChatGPT because asset classes such as
dividend or cyclical stocks are not included in classical asset allocations.

3
We interact with the publicly available version of ChatGPT (GPT-3.5).
4
ChatGPT responses are regenerated 10 times to collect three asset classes that appear most consistently. We have excluded foreign currency,
which was recommended for a single case.

3
J.H. Kim Finance Research Letters 58 (2023) 104580

Table 1
Asset class selection for every economic condition (+: rising, − : falling).
RATE INF USD Selected asset classes (ETFs)

Cond. 1 þ þ þ IYR TIP XLI

Cond. 2 þ þ – GSG IVV IXUS
Cond. 3 þ – þ DVY GSG IYR
Cond. 4 þ – – IVV IXUS IYR
Cond. 5 – þ þ IYR TIP XLI
Cond. 6 – þ – GSG IAU IYR
Cond. 7 – – þ DVY GOVT LQD
Cond. 8 – – – AGG IAU IYR

Table 2
Asset classes and corresponding ETFs.
Assets class Fund/ETF

US Treasury bonds iShares U.S. Treasury Bond ETF (GOVT)

US bonds* iShares Core U.S. Aggregate Bond ETF (AGG)
TIPS iShares TIPS Bond ETF (TIP)
Corporate bonds iShares iBoxx $ Investment Grade Corporate Bond ETF (LQD)
US stocks* iShares Core S&P 500 ETF (IVV)
International equities* iShares Core MSCI Total International Stock ETF (IXUS)
Dividend stocks iShares Select Dividend ETF (DVY)
Cyclical stocks Industrial Select Sector SPDR Fund (XLI)
Commodities* iShares S&P GSCI Commodity-Indexed Trust (GSG)
Gold iShares Gold Trust (IAU)
Real estate* iShares U.S. Real Estate ETF (IYR)
*
Major asset classes.

5.2. Backtest results for out-of-sample

We next perform out-of-sample backtest from October 2022 to July 2023, which does not overlap with the GPT training period.
Here, we simulate a realistic investment situation where the economic condition up to the current month is used to make investment
decisions for the following month. For example, information up to October 2022 is used for selecting three asset classes and the optimal
portfolio is constructed at the end of October for investment in November 2022. Similar to the in-sample tests, we use 3-, 6-, and 12-
month momentums for assessing current market conditions and use them for predicting the economic condition in the following
month.5 For portfolio construction, we construct optimal M-V, GMV, and RP portfolios with a 3- or 6-month lookback period (daily
returns) in addition to the Eqw portfolio.
Table 5 presents results for a 6-month lookback period, which shows that GPT portfolios generally have higher returns. While there
is no dominance in volatility and downside risk, GPT portfolios have higher Sharpe ratios (unless negative) compared to investment in
five major asset classes.6 Thus, there appears to be value in forming more concentrated portfolios (three ETFs) based on ChatGPT to
achieve higher gain during the market upside. Considering that GPT portfolios are only exposed to only three asset classes at any given
time, the efficiency supports the value of asset class recommendations by ChatGPT.
Finally, we further analyze the effectiveness of ChatGPT’s asset class selections by comparing with Eqw portfolios that randomly
select three asset classes each month. In particular, we use the top 30 iShares ETFs based on net asset value, which include various
funds invested in stocks, bonds, commodity, and real estate7. Fig. 3 shows a histogram of terminal wealth of 10,000 random portfolios
along with GPT portfolios when assuming a value of one at the beginning of October 2021. The terminal values of random portfolios
have a mean of 0.987 and a median of 0.985. The GPT-based Eqw portfolios are at the top 18.1 (GPT3m), 20.1 (GPT9m), 24.2
(GPT6m), and 42.2 (GPT12m) percentiles. Even though we do not search for an optimal economic prediction model, simply using
various momentums results in out-of-sample performance within the top 18 to 42 percentiles. This further demonstrates the advantage
of using ChatGPT for selecting asset classes based on economic conditions.
The key significance of our finding is that the use of ChatGPT provides a framework for incorporating economic conditions for
making investment decisions in relatively short frequencies. It is difficult to use economic indicators in a quantitative approach such as
factor models because indicators are mostly measured on a monthly or quarterly basis, which does not provide enough data points for
performing significant time-series analysis. Nonetheless, ChatGPT can act as an individual that learns market insights through various

5
We use momentum longer than one month since we need one-month predictions.
6
Negative Sharpe ratios are omitted from Table 5 because interpretation is often not intuitive (McLeod and van Vuuren, 2004). Results for a
3-month lookback are included in the online appendix.
7
The 30 ETFs are IVV, IEFA, AGG, IJH, IWF, IEMG, IJR, IWM, IWD, EFA, ITOT, TLT, IVW, MUB, QUAL, IXUS, LQD, IWB, USMV, IWR, IEF, IAU,
MBB, SHY, IVE, DGRO, GOVT, IGSB, IUSB, and TIP.

4
 J.H. Kim
Table 3
3-month momentum from 2013 to 2022 (+: rising, − : falling).
}
Mar

Aug

Oct
Jan

Apr

Nov

Dec
Sep

Mar

Aug

Oct

Nov

Dec
Jan

Apr

Sep
Feb

May

Jun

Feb

May

Jun
Jul

Jul
2013 RATE – – – þ – – – – – – – þ – – – þ þ þ þ þ – – – þ
to INF – þ – – – þ þ þ – – – þ þ – þ þ þ þ þ – – – – –
2014 USD þ þ þ þ þ þ þ þ þ – – þ þ þ þ – – – – þ þ þ þ þ

2015 RATE þ þ – þ þ þ þ þ þ – – þ þ þ þ þ – þ þ þ þ þ þ þ
to INF – – – – – þ þ þ – þ þ þ þ þ þ – þ þ – þ þ þ þ þ
2016 USD þ þ þ þ þ – þ þ þ þ þ þ þ þ – – – – þ þ þ þ þ þ
5

2017 RATE þ þ þ þ þ þ þ þ þ þ þ þ þ þ þ þ þ þ þ þ þ þ þ þ
to INF þ þ þ – – – – þ þ þ þ – þ þ þ þ þ þ þ – – – – –
2018 USD þ þ – – – – – – – – þ þ – – – – þ þ þ þ þ þ þ þ

2019 RATE þ þ þ þ – – – – – – – – – þ – – – – þ þ þ þ – –
to INF – – – þ þ – – – þ – þ þ þ þ – – – – þ þ þ þ – –
2020 USD – – – þ þ þ þ þ þ þ – – – þ þ þ þ – – – – – – –

2021 RATE – – – – – þ þ þ þ – – – – – þ þ þ þ þ þ þ þ þ þ
to INF þ þ þ þ þ þ þ þ – þ þ þ þ þ þ þ þ þ þ – – – – –
2022 USD – – þ þ – – þ þ þ þ þ þ þ þ þ þ þ þ þ þ þ þ þ –

Finance Research Letters 58 (2023) 104580

J.H. Kim Finance Research Letters 58 (2023) 104580

Fig. 2. Comparison of in-sample portfolio performance (log wealth).

Table 4
Monthly in-sample performance of Eqw portfolios.
GPT1m GPT3m GPT6m GPT12m All assets 5 assets

CAGR 0.0847 0.0741 0.0751 0.0466 0.0596 0.0561

mean 0.0072 0.0063 0.0065 0.0042 0.0051 0.0050
std 0.0281 0.0271 0.0295 0.0277 0.0228 0.0302
Sharpe ratio (rf=0) 0.256 0.234 0.220 0.151 0.223 0.166
mean (ann.) 0.0863 0.0761 0.0778 0.0502 0.0612 0.0603
std (ann.) 0.0975 0.0940 0.1020 0.0959 0.0791 0.1047
Sharpe (rf=0, ann.) 0.885 0.809 0.762 0.524 0.774 0.576
MDD − 0.1686 − 0.1382 − 0.1348 − 0.1774 − 0.1482 − 0.2221
VaR95 % − 0.0417 − 0.0416 − 0.0417 − 0.0516 − 0.0312 − 0.0403
CVaR95 % − 0.0661 − 0.0579 − 0.0591 − 0.0622 − 0.0508 − 0.0716

Table 5
Monthly out-of-sample performance of portfolios (6-month lookback).
Portfolio model Eqw Risk parity

Asset selection GPT3m GPT6m GPT12m All assets 5 assets GPT3m GPT6m GPT12m All assets 5 assets

CAGR 0.0051 − 0.0001 0.0036 0.0018 0.0031 0.0014 − 0.0009 0.0062 − 0.0035 − 0.0048
mean 0.0031 0.0006 0.0031 0.0015 0.0023 0.0012 0.0001 0.0033 − 0.0012 − 0.0017
std 0.0395 0.0366 0.0551 0.0354 0.0420 0.0332 0.0325 0.0331 0.0282 0.0311
Sharpe ratio (rf=0) 0.0783 0.0166 0.0569 0.0410 0.0548 0.0361 0.0038 0.1009 – –
mean (ann.) 0.0371 0.0073 0.0376 0.0174 0.0276 0.0144 0.0015 0.0400 − 0.0145 − 0.0203
std (ann.) 0.1367 0.1269 0.1910 0.1226 0.1455 0.1151 0.1126 0.1145 0.0978 0.1079
Sharpe (rf=0, ann.) 0.2712 0.0574 0.1970 0.1420 0.1898 0.1251 0.0130 0.3497 – –
MDD − 0.1914 − 0.1750 − 0.2215 − 0.1521 − 0.1689 − 0.1736 − 0.1578 − 0.1317 − 0.1532 − 0.1625
VaR95 % − 0.0579 − 0.0464 − 0.0882 − 0.0529 − 0.0653 − 0.0467 − 0.0406 − 0.0366 − 0.0386 − 0.0431
CVaR95 % − 0.0618 − 0.0550 − 0.0912 − 0.0640 − 0.0761 − 0.0504 − 0.0484 − 0.0463 − 0.0505 − 0.0562

Portfolio model GMV M-V 5 %

Asset selection GPT3m GPT6m GPT12m All assets 5 assets GPT3m GPT6m GPT12m All assets 5 assets

CAGR − 0.0058 − 0.0039 − 0.0019 − 0.0100 − 0.0091 − 0.0034 − 0.0036 − 0.0014 − 0.0125 − 0.0058
mean − 0.0023 − 0.0013 − 0.0004 − 0.0043 − 0.0039 − 0.0012 − 0.0012 − 0.0002 − 0.0054 − 0.0022
std 0.0279 0.0291 0.0310 0.0201 0.0219 0.0282 0.0305 0.0319 0.0265 0.0300
Sharpe ratio (rf=0) – – – – – – – – – –
mean (ann.) − 0.0271 − 0.0161 − 0.0047 − 0.0520 − 0.0468 − 0.0139 − 0.0143 − 0.0019 − 0.0642 − 0.0262
std (ann.) 0.0965 0.1008 0.1075 0.0696 0.0759 0.0977 0.1055 0.1106 0.0917 0.1038
Sharpe (rf=0, ann.) – – – – – – – – – –
MDD − 0.1713 − 0.1630 − 0.1645 − 0.1389 − 0.1456 − 0.1431 − 0.1603 − 0.1587 − 0.1582 − 0.1570
VaR95 % − 0.0401 − 0.0363 − 0.0363 − 0.0358 − 0.0382 − 0.0402 − 0.0435 − 0.0435 − 0.0377 − 0.0353
CVaR95 % − 0.0401 − 0.0399 − 0.0399 − 0.0399 − 0.0422 − 0.0404 − 0.0461 − 0.0461 − 0.0605 − 0.0543

sources that may include views of experts and consensus of market participants.

6. Conclusion

This study proposes how a large language model such as ChatGPT can be incorporated into a quantitative investment model. Since
ChatGPT can learn macro relation between economy and financial markets, we use ChatGPT as a quant asset manager that recom­
mends asset classes given economic conditions. Our empirical findings show that asset class recommendations by ChatGPT improves

6
J.H. Kim Finance Research Letters 58 (2023) 104580

Fig. 3. Terminal wealth of GPT and random portfolios.

portfolio efficiency compared to more diversified portfolios. Given that these efficiencies are observed with the use of only three
indicators and three asset classes, we believe there is opportunity for quantitative investment approaches to benefit from utilizing large
language models. Interestingly, the three indicators chosen in our analysis are market indicators that are also published with daily
frequency (e.g., breakeven inflation rate measures daily expectations). While our results show significance when performing analysis
on a monthly basis, further research with higher frequency could provide more insight into using large language models for market
analysis and investment management.

CRediT authorship contribution statement

Jang Ho Kim: Conceptualization, Methodology, Software, Writing – original draft, Writing – review & editing.

Data availability

All data used are publicly available (FRED and Yahoo Finance).

Acknowledgments

This research was supported by the BK21 FOUR (Fostering Outstanding Universities for Research) funded by the Ministry of Ed­
ucation (MOE, Korea) and National Research Foundation of Korea (NRF), and partly supported by Institute of Information & com­
munications Technology Planning & Evaluation (IITP) grant funded by the Korea government (MSIT) (No.RS-2022-00155911,
Artificial Intelligence Convergence Innovation Human Resources Development (Kyung Hee University)).

Supplementary materials

Supplementary data associated with this article can be found, in the online version, at 10.1016/j.frl.2023.104580.

References

Aldridge, I., 2023. The AI revolution: from linear regression to ChatGPT and beyond and how it all connects to finance. J. Portf. Manag. 49 (9), 64–77.
Czasonis, M., Kritzman, M., Pamir, B., Turkington, D., 2020. Enhanced scenario analysis. J. Portf. Manag. 46 (4), 69–79.
Dowling, M., Lucey, B., 2023. ChatGPT for (finance) research: the Bananarama conjecture. Financ. Res. Lett. 53, 103662.
Gavell, S., Kritzman, M., Kulasekaran, C., 2022. Severe but plausible – or not? J. Risk 24 (3), 97–119.
Greenberg, D., Babu, A., Ang, A., 2016. Factors to assets: mapping factor exposures to asset allocations. J. Portf. Manag. 42 (5), 18–27.
Kim, J.H., Lee, Y., Kim, W.C., Fabozzi, F.J., 2021. Mean–variance optimization for asset allocation. J. Portf. Manag. 47 (5), 24–40.
Ko, H., & Lee, J. (2023). Can ChatGPT improve investment decision? From a portfolio management perspective. Available at SSRN: 10.2139/ssrn.4390529.
Kritzman, M., Li, D., Qiu, G., Turkington, D., 2021. Portfolio choice with path-dependent scenarios. Financ. Anal. J. 77 (1), 90–100.
Lo, A.W., Singh, M., ChatGPT, 2023. From ELIZA to ChatGPT: the evolution of natural language processing and financial applications. J. Portf. Manag. 49 (7),
201–235.
McLeod, W., van Vuuren, G., 2004. Interpreting the Sharpe ratio when excess returns are negative. Invest. Anal. J. 33 (59), 15–20.
Podolsky, P., Johnson, R., & Jennings, O. (2012). The all weather story. Bridgewater Associates, White paper.
Qian, E., 2011. Risk parity and diversification. J. Invest. 20 (1), 119–127.
Romanko, O., Narayan, A., & Kwon, R.H. (2023). ChatGPT-based investment portfolio selection. aRxiv:2308.06260.
Saggu, A., Ante, L., 2023. The influence of ChatGPT on artificial intelligence related crypto assets: evidence from a synthetic control analysis. Financ. Res. Lett. 55 (B),
103993.