Skip to content
All library documents

Limits and Approaches to Inferring Trading Strategies From Trade Data

Article Quant Q&A · Author: Dmitri Nesteruk
Read the original

Summary

This note considers whether a participant’s algorithm can be reconstructed from observed competition trades. The responses emphasize that trade streams alone leave many possible explanations because strategy decisions depend on hidden variables, market conditions, and randomness. One suggested alternative is to analyze market bars as hypothetical directional trades, though this is only a simplified way to study market behavior and does not reproduce a participant’s actual decision process.

The note also describes a research approach that selected high-performing trade streams as training examples for neural-network pattern discovery, but reports no outcome and leaves its effectiveness uncertain. A cited simulation study proposes identifying the best-performing agent and examining its operations in a structured trading space, including a lookup-table mechanism. These examples are exploratory; the document provides no validated inference method, performance evidence, or practical guidance for obtaining suitable data.

Key ideas

  • Observed trades generally do not uniquely reveal the algorithm that produced them.
  • Market bars can be treated as simplified hypothetical long or short trades for analysis.
  • One proposed research approach trains neural networks on selected high-performing trade streams.
  • A cited simulation study analyzes the operations of the best-performing agent in a trading space.
  • The described approaches have no reported success evidence, and data access may raise legal concerns.

Tags

Machine learningStatisticsBacktestingMarket microstructure
Full text
# Means of inferring trading algorithms from competition trade data


# Means of inferring trading algorithms from competition trade data












I'm analyzing trades from several participants in a trading competition, and I was wondering - are there known mechanisms for analysis and inference of the logic in a set of trades done by one participant? For example, I know that some participants would adopt the strategy of hedging an option-on-future O with the underlying future F, but I have no idea what greeks they are hedging and what other states in the system their strategy might be taking into account.

## Answer by alpha (score 4, accepted)

https://quant.stackexchange.com/a/3170

Mathematically speaking that is an impossible thing to do. There are simply too many variables and randomness that you cannot do it.

Rather than analyzing competitors trade data; why don't you analyze the market.

You can consider every bar; as a trade. If the bar was up; it means you went long and made a profit. If the bar was down; it means you went short and made a profit.

## Answer by Samik R (score 2)

https://quant.stackexchange.com/a/3181

Agree with the previous post, that it is very difficult to do, if not impossible. Having said that, I once saw a presentation on a (closed) project which was trying to use neural networks to train and form trading patterns. The group had access to a large lot of trading data and the profit (or losses) made from the trades over a certain time period (like, a year or so of data). They filtered that data to choose top few thousands trade streams (top as measured by profit at the end of the time period), and then used that as the training set.

I do not know the end result, so can't say if it was a successful project. Not to mention, I think, there are potential legal hazard in acquiring such trading streams, if they are available at all. But the idea is worth mentioning nevertheless.

## Answer by Ram Ahluwalia (score 2)

https://quant.stackexchange.com/a/3203

This paper by R. Marschinski and H. Kantz may be able to help: "Learning the Optimal Trading Strategy". I have not read it but the authors have published other work which is quite innovative.

> Within a realistic model of the stockmarket, we derive the most successful trading strategy. We first identify the agent who has realized the largest percentual gain and then analyze all the operations this trader has performed during the simulation run. We report them in a proper trading space and we extend the model, introducing an additional operator acting with the help of a look up table derived from a clusterization of space. We discuss the robustness of this optimal strategy, its performance and the applicability to real markets.

Shown in full with attribution under the source's licence. Licence: CC BY-SA 4.0 (Stack Exchange)

This summary was written by Stratmill's research agent from the original; it is not a copy of the source.