I think the 109 values you call "tests" are usually called the validation set, but this is a matter of nomenclature.

In any case, I am surprised this can work at all. I've always thought market prices are more or less random walks. In fact, I'm skeptical about the very possibility to predict them from a theoretical point of view, based on the following reasoning.

What would happen if such models turned out to be really efficient at predicting the price and if such model was made public? Wouldn't everyone eventually use it, making the prediction self-fulfilling?

In fact, there could be a positive feedback happening. The model would predict prices fluctuations that would be amplified by speculators. In the end the model itself would direct the market and the price would become a loose degree of freedom.

He's predicting log-price levels, not returns (log-price differences). It is the returns that are random, not the prices. The price series themselves are highly autocorrelated so you could get similarly good-looking results with the naïve forecast function LogPrice(t+1) = LogPrice(t).

The real test would be to use the models to try to predict returns (by differencing the forecast log-prices). But the results would be disappointing, just as they would be if you had used, say, an ARIMA model to predict the prices. The returns, which are actually what you are interested in, as just unforecastable "residuals" as far as such price models are concerned.