Statistical Algorithms vs. Active Monitoring Algo's

[Harkkam]

Hey guys I had a quick question, forgive me if its basic or seems silly. But after doing some research I came across a type of algo's that take past data and try to construct models around this past data. If you have the past 5 years of returns on a stock, the algo is designed to create a model of a normal pattern a roadmap of the stock over these past 5 years and whenever the stock deviates from this normal pattern it executes a trade in the opposite direction betting for a return to normal. For example if every time earnings reports come out the stock has fallen 2% 3 times out of 5 then an algo positions itself to short the stock the second earnings come out

However this type of algo is a little different that doesnt take past data into account to develop the algorithm. It actively monitors indicators like MACD, vol, 200 day avg, ketner channels etc. This algo has a complex chain of requirements that a stock must fit like having earnings report on vol greater than 2 million with massive deviation from its vwap with an ATR of 5. The algo is constantly scanning the stock universe for stocks that fit its set parameters. However its not using what the stock did in the past to base its decisions to buy or sell, its not using historical prices at all.

In my mind the first type of algo suffers from a look back bias, I work as a prop trader right now and one of the things I notice is that institutions move prices and that supply and demand of a stock can create massive moves one year and not the next.

You can see on the Level 2 when a large buyer appears and keeps bidding the stock up and up. This action this participation by this buyer is unique to this event, a hedge fund wants to invest and causes a 3 point move over the day.

The past looking algo cannot predict future supply and demand pressure from just looking at past data, the value of a stock is constantly changing with its perception.

My question is on what basis do past looking algo's have validation, when stocks move based on news that cannot be predicted from past data alone

 

[Yike Lu]

Good technique for a backtest includes holding some of the data back for "validation", in other words you run it "forward" on a simulated future (neither you nor the algo have seen the data before). For additional safety you can simulate on market data as it comes in.

For the second type of algo, how did somebody come up with it? Obviously one has to use observations about the past. It may not actively look in the past, but somebody looked at past data (whether backtesting or "live" experiencing the trading firsthand) and determined that it would be profitable to trade under certain conditions.

For example, how did you come to the conclusion that institutions move markets, and that this is detectable in Level 2? You watched it, and so used past data to draw conclusions about the future.

To me it is 12 of one, a dozen of the other. Same thing, but variations. Philosophically there's relatively little difference, although the devil is in the details.