Quick peak into more of the analysis of our next-gen trading model

== LEFT CHART - STRATEGY GENERALIZATION PLOT: ==

This is the most important plot

This shows the ability of our top-performing carefully-engineered objective function to rank strategies based on their performance on e.g. 100 days of seen data, so as to maximize performance on the subsequent 100 days of unseen data

Ideally, though generally impossible, would be this point cloud exactly matching the diagonal line, e.g. 100x on the seen data -> 100x on the unseen data

Worst-case, you want to outperform the horizontal line, e.g. 100x on the first half -> 1x on the second half

As you can see, our new method excels incredibly, achieving a point in the middle, while balancing well for things like drawdown duration: 100x on the first half -> 10x on the second half. = Surprisingly good generalization ability,

And you can see all the quality measures sloping upward as you approach the top-ranked strategies. Very hard to achieve.

== RIGHT CHART: ==

Here’s more details on the performance of an ensemble of those top strategies from the left chart looks like

Through careful ensembling methods, you can see it achieving an incredible ~80% of the days higher than the previous days, and a compounding ROI of 60x over this duration, with quite smooth exposure over time

Remember that it’s quite a huge challenge to achieve performance with this kind of “skew” — of simultaneously achieving (1) drawdown of just 20% of the days, while still (2) achieving high ROI

— Usually these 2 things are in strong conflict, fairly easy to get either one while totally trashing the other, very hard to get both

But that’s what we achieved, and now set to launch this new model

For comparison, our current model achieved strong ROI, but was much weaker on the preventing of drawdowns.

This new approach achieves both, thanks to big advancements in:

(1) Better objective functions
(2) Better ensembling methods
(3) Better AI guided search of the strategy space, vastly speeding up searches that’d otherwise take millions of years to brute force

Lots more to the math, but these are the most important concepts:

(1) Very strong generalization performance on a proper future holdout set — Left chart
(2) Good characteristics of the selected ensemble — Right chart
(3) Excellent performance in live forward testing, matching the backtesting shockingly well

^^ There’s a brief glimpse into the math going on behind the scenes with a real, proper AI memecoin trading system — AFAIK the first of its kind?

🄳🄾🄾🄼🄿🄾🅂🅃🄸🄽🄶

Rich Russians Hunting Somali Pirates on Luxury Yachts (2009)

At the height of the Somali piracy crisis in 2009, when hijackings had turned the Gulf of Aden into one of the world’s most dangerous shipping lanes, one Russian security company proposed an extraordinary business model:

Instead of avoiding pirates, wealthy tourists would go looking for them.

For around £3,500 per day, clients boarded armed luxury yachts departing from Djibouti and sailed deliberately close to the Somali coast.

The yachts were reportedly staffed by former Russian special forces and equipped with military-grade weapons ranging from AK-pattern rifles to grenade launchers and rocket launchers.

Their role was simple:

If pirates attempted to board, the passengers could participate in defending the vessel under the supervision of experienced professionals.

The week-long voyages followed the route from Djibouti to Mombasa, Kenya, passing directly through waters where hundreds of merchant ships had been attac