Multisided Fraction Executions in skNQx
What we're working on in 2024-2025
Observing the following Feedback Loop: (A ≠ B), in skNQx, allows for another possibility. Opening trades in both directions, s & k. ShezheaNET MLT_HFT is now not dependent on directional microsecond-TE but can act as a multi-sided hedging function.
Multisided Fraction Executions are currently not possible but will be enabled with streamlined trading getting more efficient. We expect this practically tested algorithm within ShezheaNET MLT_HFT to be operational under real-time market conditions within a year.
Considering the parameters-concept (q/a/y/r/d) in oNQ9 MLT-HFT, it wouldn't make sense to use said parameters for multi-directional executions, since A has to equal B while -A equals B as well after a chronical dimension. The target is thus to find (A) that leads to (q/a/y/r/d), but at the same time (-A) that leads to (q/a/y/r/d) after a specific duration. This is what ShezheaNET calls Multisided Fraction Executions. In simpler terms: Longing and Shorting a Market at the same time ("hedging"), while making the maximum amount of capital in both directions in seconds. Note that this is only possible in the "Optimal 1/9 SeekDestroy NQ Model" (Sub-Model A$) -> skNQ9. To further dissect the skNQ9 TE, the ShezheaNETs algorithm uses the following logic: skNQ9 is designed to be multidirectional, offering liquidity in both directions and filling larger orders. We observe (here simplified, in the practical code it's described as TimestampQ, and further in sections 1/2/3, and TimestampR, and further in sections 1/2) different market conditions between s & k. This is a critical issue for our 20x3 data cluster execution model. s now ends up with a (q/a/y/r/d) result that is impractical for k, thus resulting in a feedback oxymoron and an overall loss in PNL. The output (A) is correct, since A=B, but (-A) cannot ever be close to (B). This issue is critical because it lowers the possible PNL by -70% - -150%. The option to just "splitter" the data clusters is inefficient as well since the outputs are reliant on each other. The target is to find (A) ≠ (B) until f(q, a, y, r, d) is maximized, while also having (-A) ≠ (B) until f(q, a, y, r, d) is maximized. This has to happen in an active/live TE, in a time window of max. 2 sec, and to have a feedback loop that is independent of each other.
Understanding the significance of the mentioned issue is crucial. 83% of TEs in NQ are in either skNQ9 or bskNQ9, meaning finding a solution for this, would allow us to increase HFT Activity by over +700%, and HFT profitability by over (est.) +500%.
We now consider the following two options in the post-execution phase. Option 1: (A); (-A) & Option 2: (-A); (A)
We found the following divergences between these options:
(A); (-A) is more probable than (-A); (A) if expected y > 0.1314296
Root Mean Square Error (RMSE) is fairly higher in (-A); (A)
(-A); (A) can be viewed as a fixed hedged trade that gets opened with a delay to its counter-position.
Last updated