Triple Barrier Labels Represent One Outcome per Trading Instance

Summary

This exchange clarifies how to interpret labels in the triple barrier method. Each labeled instance corresponds to a specific trading circumstance and receives a single outcome according to which event closes it: an upper barrier win, a lower barrier loss, or a time-based close. The answer says the label represents that instance rather than creating a new observation at every step along its path.

The response addresses whether observations should be generated at every time step or only when a barrier is touched. It emphasizes that the horizon's elapsed time does not change the basic unit of labeling: one instance receives one outcome. The exchange is brief and does not explain how to choose barrier levels, define event start times, handle overlapping holding periods across separately initiated instances, or encode labels for a particular strategy. Those implementation details require a fuller specification of the trading domain.

Key ideas

• Each trading instance receives one label based on the event that closes it.
• The possible outcomes are an upper barrier win, a lower barrier loss, or a time-based close.
• The label describes the instance rather than assigning a separate outcome at every time step.
• Barrier selection and overlap among separately initiated instances are not detailed in the answer.

Tags

# Answer by Paralosva Rosos (score 1, accepted)


# Lopez de Prado's Triple Barrier Method - do you reset the barriers on every timestep, or only when you reach a hi/lo/no barrier hit?












I've used de Prado's trend scanning labels but now I want to try the triple barrier method idea.

I understand the rough set up, and I understand the logic behind it. However, what I don't understand is this: Do I have an observation on every single timestep, with the label being -1/0/1 depending on which barrier is hit first? OR do I only have a new observation at the point at which a barrier was hit (so if +1 barrier was hit at t=5, my 2nd observation starts at t=5)?

Essentially, do I have "overlapping" labels or not?

I hope my question makes sense. Thank you!

## Answer by Paralosva Rosos (score 1, accepted)

https://quant.stackexchange.com/a/73430

No, it's only one observation. Each instance is one label. You would have a specific circumstance based on the trading domain, and you would either win, lose, or close. The time would not matter.

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.