1 of 22

Synthetic Data & Artificial Stock Market

Dmytro Pokidin, Adaptive Investing Equity

Q3 2022

Format the title text in three colours according to the house style guide: use the macro in the NN Text formatting menu

Do not change or add other logos

2 of 22

Agenda

  • AI Equity Artificial Stock Market
  • Applications

2

3 of 22

AI Equity Artificial Stock Market

4 of 22

Artificial Stock Market

The challenge

  • Typically, investment strategy is based on a back-test using real historical data
  • However, back test overfitting often occurs: an investment strategy or a quantitative model is selected based on good historical in-sample performance, yet it fails to generalize well out-of-sample
  • Growing evidence of academic papers based on spurious factors and failed active investment strategies

What can be done

  • Lab experiments, out-of-sample testing. Not available in finance
  • Instead: multiple testing frameworks, Bayesian priors, regularization, hold-out samples. Limited success

Our approach

  • Enable lab experiments, out-of-sample testing – Artificial Stock Market and synthetic data

​

​

Motivation

4

5 of 22

Artificial Stock Market

  • An artificial environment (agent-based model) that generates synthetic stock market data

​

  • Designed bottom-up, replicating the behaviour of main agents in the stock market with realistic behavioural assumptions

​

Intro

5

Allocators

Funds

Stocks

Shares

Price update

Fund Shares

Model participants:

    • Allocators, e.g., pension funds, insurance companies
    • Funds, e.g., asset managers
    • Companies

​

​

​

Closed system (no cash leakage)

Daily frequency

Exogenous monetary policy reflected by stochastic cash growth

6 of 22

Artificial Stock Market

Cycle of agent interaction:

  • Allocators make their investment decisions
  • Funds observe the flows from allocators and generate trade orders
  • Prices are adjusted
  • Company shares get allocated to funds
  • Funds allocate their units to allocators
  • Repeat…

​

Big picture

6

Allocators

Funds

Stocks

Shares

Price update

Fund Shares

7 of 22

Individual vs Aggregate Behavior

  • Agent behaviours were modelled individually. We then observe the aggregate behaviour of the ecosystem.

​

  • Findings:
      • Volatility clustering
      • Gain/loss asymmetry and fat tails
      • Financial cycle, boom & bust
      • No long-term autocorrelation
      • Stocks co-movement
      • Excess volatility
      • Negative skewness of market return vs. zero skewness of individual stock returns on average

​

​

While not specifically modelled, the aggregate behavior display characteristics of real stock markets

​

7

Use ‘formatting text-styles’ in the NN PowerPoint-menu

Do not change or add other logos

8 of 22

Data Characteristics from Artificial Stock Market

  • Artificial Stock Market (ASM) closely mimics the return moments observed in real markets

​

  • High kurtosis or fat tails.

​

  • Negative skewness often called a gain/loss asymmetry. The probability of an extremely high negative return is higher than a positive one

​

  • Curiously, negative skewness is observed in market returns; it is not observed (on average) in individual stocks.

​

Gain/Loss asymmetry and fat tails

8

9 of 22

Data Characteristics from Artificial Stock Market

Both return series, synthetic and real, feature near zero long run autocorrelation, measured on a 60-month rolling window

​

Return Autocorrelation

9

Real market returns refer to Kenneth French market returns (Kenneth R. French - Data Library (dartmouth.edu)), as the data is standard open-source data for research, having a longer history available than any current stock market indices.

​

10 of 22

Data Characteristics from Artificial Stock Market

Positive autocorrelation of squared returns (volatility clustering) is observed in the synthetic data generated by the Artificial Stock Market (ASM), consistent with the observation in real stock market

​

Volatility Clustering

10

Real market returns refer to Kenneth French market returns (Kenneth R. French - Data Library (dartmouth.edu)), as the data is standard open-source data for research, having a longer history available than any current stock market indices.

​

11 of 22

Data Characteristics from Artificial Stock Market

  • Artificial Stock Market (ASM) witnesses financial cycles as in real life

​

  • The financial cycles and the asymmetry of gain/loss in the artificial stock market are endogenously generated

​

Boom and Bust

11

12 of 22

Data Characteristics from Artificial Stock Market

  • Artificial Stock Market (ASM) generates returns that are more volatile than what fundamentals predict, excess volatility

​

  • The blue line is the ratio between stocks volatility and the volatility of their fundamental values*

​

  •  Excess volatility increases during turbulent times (grey shaded area: increased volatility)

​

Excess volatility

12

*Fundamental value of stocks follows a Geometric Brownian Motion stochastic process with idiosyncratic and cyclical components�

13 of 22

Data Characteristics from Artificial Stock Market

In both artificial and real stock market, we observe the tendency of an increase in stock co-movement during periods of increased volatility (grey shaded areas)

​

​

Stock co-movement

13

14 of 22

Insights from Artificial Stock Market

​

​

​

​

​

​

​

​

  • Low Volatility, Value and Size earn premia most of the time
  • Size outperformance shall be taken with a grain of salt. It is likely a result of missing mechanisms in the market (no short selling, no corp behaviour) causing market friction
  • Momentum works half of the time

​

​

​

​

​

​

​

​

​

​

Factors premia distribution

14

​

Momentm

LowVol

Value

Size

mean

-1.4%

3.3%

1.6%

7.0%

std

5.3%

3.0%

2.5%

2.3%

min

-11.4%

-15.0%

-6.2%

1.0%

5%

-7.0%

-1.3%

-1.6%

4.1%

25%

-4.5%

1.7%

0.0%

5.5%

50%

-2.6%

3.3%

1.2%

6.7%

75%

0.2%

5.0%

2.9%

8.1%

95%

8.3%

7.5%

5.9%

11.0%

max

35.5%

17.2%

27.8%

35.2%

15 of 22

Insights from Artificial Stock Market

  • Investment styles are dynamic, they come in and out of favour all the time

​

​

​

​

​

​

​

​

​

​

​

Style Rotation

15

16 of 22

Applications

17 of 22

Synthetic Data

  • Where does it shine?
      • Real data is limited/difficult to acquire
      • Tail events are important but not sufficiently represented in real data
      • Successful applications in AI (computer vision, robotics)

17

Intro

18 of 22

Synthetic Data

18

  • Due to growing evidence of success, synthetic data vendors have proliferated in the past few years

Intro

19 of 22

Synthetic Data

  • Despite proliferation of synthetic data vendors in the past years, very few, if at all, is aimed at financial markets and investing

​

  • It is hard to create synthetic data that is usable for investment strategies

​

  • Stock markets are manifestation of social laws rather than physical laws

​

19

In investing

20 of 22

Synthetic Data

  • Due to a huge promise, there is a growing interest in both industry and academia

​

  • J.P. Morgan AI Research invests in synthetic data research and agent-based simulations

​

  •  University of Oxford Computer Science recently received 3.6 mln pounds of funding for the finance-centered agent-based modelling project, they are developing an ASM as well

​

  • Refinitiv - agent-based synthetic trading data research

​

​

20

In investing

21 of 22

Application # 1

  • Validate/test quant models (alpha, portfolio construction, trading) in an artificial environment
  • Mitigate model selection bias and overfitting
  • Test hypotheses
  • Develop new investment strategies
  • Academic interest
  • Note that the ASM can be utilized by the teams that innovate around the model rather than the input

21

Synthetic data for testing

22 of 22

Application # 2

  • Develop and train deep pricing or trading models

​

​

22

Train

Neural net

Real data

Fine-tune

Deploy

Synthetic data

Neural net

Neural net

  1. New investment signals/strategies
  2. New innovative products

Alpha based on AI