Inferring Trader Profitability from Incomplete ECN Trade Records
Summary
The document considers how to assess traders when ECN records reveal only a fraction of their executions and omit their activity elsewhere. Since the missing trades prevent direct reconstruction of positions and profit and loss, the answer recommends treating the observed history as incomplete data rather than assuming it represents each trader’s full book.
The suggested method is to identify complete sequences in the available records, estimate the joint distribution of relevant variables, and use a Bayesian approach to infer missing observations. Transaction-specific constraints should inform the model, including whether short selling is possible and the likely risk, profit and loss, and time needed to unwind positions. The analysis should be repeated over rolling windows so that complete sequences can be detected as new data accumulates. These steps offer a framework rather than a validated profitability estimator; results remain dependent on the missing-data assumptions and the representativeness of the observed venue flow.
Key ideas
- Partial venue records do not reveal a trader's complete positions or profit and loss.
- Bayesian missing-data methods can infer unobserved trades from modeled relationships among variables.
- Complete observed sequences can help estimate patterns before filling identified gaps.
- Market rules, including short-sale constraints, should shape transaction and position assumptions.
- Rolling analysis can capture complete sequences as the observed trade history develops.
Tags
Full text
# Analyzing an incomplete set of trades # Analyzing an incomplete set of trades Say I have access to logs of all trades executed on an ECN with the price maker and taker named. These traders are only executing some of their flow on this ECN (anywhere from 5%-80% of their volume). I do not have any information about their trades done elsewhere, so I can not say for sure what their positions are and therefore what their P&L is. But I can still get an idea of who is buying low and selling high. Is there any standard approach or literature on how to analyze the profitability of an incomplete set of trades like this? ## Answer by lehalle (score 2, accepted) https://quant.stackexchange.com/a/3904 There is no standard way in quant finance to do this. Nevertheless you can use: - standard ways in statistics to deal with incomplete data sets; - specific points to take into account that you are dealing with transactions. The standard statistical way to manage incomplete data is to use a Bayesian method: you model the usual cross-dependences between the variables and replace missing points by the max likely one (like in Inference and missing data by Donald B. Rubin). You can add some specific considerations like the fact that short selling is probably not allowed (except if you are working on a market where it is possible). more generally, it means that you have to compute some characteristics of what you see from the strategies, like the PnL, the risk, the time to unwind positions, etc. You should do it a sliding way (i.e. over a sliding window of few days / weeks), so that you would be able to detect some full sequences you have. Practically, it means that you need: - to isolate some full sequences you have - infer the joint distribution of all your variables - use it to fill gaps that you will have identified in your dataset
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.