[Project 88 Part 9] 139% AAPL gain?

[Project 88 Part 9] 139% AAPL gain?


I want to share with you some prediction results we’re seeing with Project 88 (You can catch up here with Part 1).

The statistical significance is on the low side, but the results with AAPL show an annualized gain of 139% by holding on days that are predicted to go up.

Now that number makes a lot of assumptions like friction-free trading on the close, but it’s a start in the right direction.

Actually, it’s 298% compounded instead of summed, but that’s a ridiculously high number devoid of any reality.

The days that are predicted to go up gain 0.55% on average.

The days predicted to go down go UP 0.32%.

Unfortunately they don’t go down, but those days don’t go up much either.

I would very much like to have some great downside predictions because shorting stocks would really give us a tremendous boost in return…and smooth out drawdowns even more.

I feel like a proud papa.

It’s like watching a child grow from an infant to an adolescent…the diaper changing, first steps, scraped knees, riding a bike with training wheels …all within a few months.

We’re getting closer and closer to adulthood.

I’m sure there will be fits and starts, gut wrenching setbacks, and shouts of joy as we continue.

The doors a strategy like this could open…well, it's a heck of a dangling carrot.

Now you might be wondering why the heck we only back tested about nine months of just Apple.

Why not 10 years? Why not the entire S&P 500?

Well, we have to get the code working on a single computer first before we replicate Project 88 in the cloud with 2000 cores at its disposal.

A bug on 2000 cores wastes a HUGE amount of money.

However, it takes over 30 minutes just to split out a few months of results on Ruud’s rather fast XEON powered computer.

Remember, a single day’s prediction has to match the current pattern to 30 years of history on 50,000 symbols. It’s a computing monster!

It’s even worse now because we relaxed a few parameters to get more pattern matches.

Only 44% of days had a prediction with statistical significance…so there’s an important tradeoff between how exact we match the pattern in a stock, and how many matches we get.

If you only have a few pattern matches, then the prediction is unreliable.

One of the ways we validated Project 88 was working correctly was to see if a low number of matches ruined predictions.

They certainly do! Accuracy falls dramatically.

By the way, no pattern ever exactly matches another because there’s a lot of noise in price data.

It’s like back in the day when the rabbit ears on your old CRT TV would need to be adjusted.

You could sort of make out what was going on in the TV show, but it sure wasn’t ideal right?

So you (if you were the youngest person in the room) would have to start grabbing the rabbit ears and tilt and cajole them until the picture cleared up into something watchable.

Man, I miss those days.

Not dicking around tuning in reruns of M*A*S*H…more like the simplicity of the era.

Anyway, we’re testing pattern matching settings that generate around 10x as many matches, but it completely bogs down the computer to the point of being ridiculous.

So that’s where we’re currently at with Project 88…making the latest code work with our 2000-core supercomputer.

So a year's worth of predictions can be done in around 3.6 seconds.

At that rate, we can make ten years of predictions on the entire S&P 500 in about 5 hours.

Where this gets really cool is when we only trade the cream of the crop stocks predicted to go up the most in a short period of time.

You can get even fancier and trade only the top stocks adjusted for both reward and risk (using something like profit factor on the price predictions).

The way I see it, to get Project 88 working, we need to:

–          Code the pattern matching/prediction algo into distributed computing code.
–          Analyze, tweak, and finalize code.
–          Add 10 years of predictions to the entire stock database.
–          Automate daily prediction updates.
–          Modify Portfolio Boss to work with predictions (database hookup, new buy/sell filters, ranking functions)
–          Work closely with Unite members to quickly generate new trading strategies and ideas for new filters and functions.

I think we’re about 70-80% there. Building the supercomputer was the really hard part, but our lead programmer is pretty amazing, and gets things done.

OK, I’m probably getting ahead of myself here, but I’m super excited to share this with you!

More from me soon!

Oh…and here's a touch of M*A*S*H nostalgia.


Trade smart,