Yes! You're right, I used multiple regression models on hourly bars for a while several months ago and the results were fairly random I think. I did tweak it as you guessed, and I'm using it in conjunction with some other tools too; multiple moving averages and simple bar reversal patterns.
The way I tweaked the MRA model was to look ONLY at bars in the London session. I group the first 5 hours together to see how the session is going overall. Then I look at the sixth hour by itself because the sixth hour is when most of the economic releases come out in the U.S. The specific inputs I use are:
the distance from London open to high of the first 5 bars,
the distance from London open to low of the first 5 bars,
the distance from London open to close of the fifth bar,
open to high of sixth bar
open to low of sixth bar
open to close of sixth bar
So with these six inputs I get a picture of how the London session was going before the news out of New York, and then how the market reacted to the U.S. news and the New York open. By that time, there are three more hours of the London session to go. The model projects a high low and close of the next hour and the remainder of the whole session.
Once I have that projection, I use it as a guide. I watch ultra low TF charts to see when the price deviates from the projection and wait for a big bubble in the fast MAs and a reversal pattern.
I've only been doing it for two days, and got a few pips both days. However, the projections have been pretty close both days, and I missed out on a big move today because I didn't bother waiting for the last two hours of the London session. I looked later, and the price had gone way above my upper projection, turned around, and hit my lower projection right around 1/2 hour before London close.
I almost hurled, LOL! That would have been a nice trade.