The Dilemma of Adaptive Forecasts

Today I am once again starting to smash one of the hypotheses to pieces.

I had an idea to look at the market as waves that move from a low to a high and back, and to measure these segments across several timeframes at once. This is not Elliott Wave Theory, but a slightly different logic.

After several backtests, I again come to the conclusion that it gave practically no competitive advantage. It highlighted a couple of things that give only small hints for further development and research.

I am already into the third hundred of different price forecasting hypotheses, under different sauces. And every time everything runs into some kind of chaos. Forecast accuracy starts to look like a phantom that you constantly have to chase.

But some work remains and accumulates. Week after week, month after month.

Probably the main concept I have reached is the following. Any algorithm, before giving a market forecast, should assess several key elements. And these are not individual metrics that can be obtained from an exchange through an API, but tendencies.

Traders call this a multi-timeframe tendency. We see that the market has been in state X for a long time. On a smaller timeframe, state Y appears. Our task is to calculate the nearest state Z.

I increasingly come to the conclusion that a fixed risk calculation formula will not work. It has to be recalculated almost every time, using the same standards a little differently.

Here a dilemma appears. Adaptability is the right move. The market is never the same: conditions, circumstances, causes and effects change. Similar patterns exist, but the context, seasonality and liquidity can be different. Too many points have to coincide, and this practically never happens.

It turns out that a forecasting system should adapt to the market and the current situation. But adaptability means that there is nothing completely permanent and stable inside it.

And here appears the problem with the backtests that everyone wants to see: “Look, my system showed an advantage and a win rate over a certain period.” The approach to such a backtest itself starts to contradict adaptability. It becomes critically difficult to conduct it honestly.

Especially if the adaptation is performed by an LLM. The model may remember what the Bitcoin price was on a specific day in 2023, 2025 or 2026. It is impossible to make it forget this.

When we ask such a model to adapt rules to historical data, it may unconsciously use knowledge from memory and fit the mechanism to a market it has already seen. In the end, we get a very beautiful, sweet, but self-deceiving picture.

The market changes all the time. Rules should be born almost here and now and be used a little differently each time. The tool that can perform such adaptation is artificial intelligence. But the same tool may already know the answers inside the historical test.

This is the dilemma I am working on now.