We're building an adaptive prediction engine. Could the same approach improve algorithmic trading?
Hi everyone,
We're developing Nexyria, an independent quantitative intelligence project focused on forecasting, market analysis and evidence-based decision systems.
One of our research projects is APEX, an adaptive prediction engine.
The idea behind APEX is relatively simple:
Instead of assuming that one forecasting method works best in every situation, we investigate whether a system can identify the current data environment and select an appropriate method.
We've started testing this approach on business forecasting datasets.
Some early results:
• APEX Business V0.1 improved RMSLE by approximately 1.13% over its baseline in an external Kaggle evaluation.
• APEX Business V0.2 was evaluated on 1,885 daily observations from the Walmart M5 dataset.
• In that second evaluation, APEX matched the strongest baseline but did not outperform it.
• The result was classified as MIXED under our predefined validation protocol.
This second result was particularly interesting.
APEX selected the same forecasting method across all 1,885 observations.
It performed reasonably well, but it did not demonstrate meaningful adaptive behavior.
That raises an important research question:
How can we distinguish genuine adaptation from a selector that simply learns to favor one strong baseline?
We're now considering whether this research approach could eventually be useful in our separate algorithmic trading system, Nexyria Trader.
For example:
Could an adaptive selector help identify different market regimes?
Could it determine when certain trading strategies are more appropriate?
Could it identify environments where the safest decision is not to trade?
We haven't validated APEX for trading, and we're not claiming that it improves trading performance.
Our intention would be to research it offline first, using historical data, walk-forward testing and forward validation, before considering any integration.
I'm interested in hearing from people working on adaptive systems, regime detection or quantitative strategy selection.
What approaches have you found most useful for distinguishing real adaptive improvements from overfitting?
And what validation methods would you consider essential before trusting such a selector in a live trading environment?