Algo Deployment by Users Class

	Dark Pools Plunge Protection Teams Sovereign Funds	Commercial Banks Electric Power Producion & Distribution Companies Oil&Gas	Treasury Engines Liquitidy Engines Clearing houses	Invest. Banks Hedge Funds	Asset Management Companies Wealth management Companies Family Offices
User Acceptance Testing	Specific	Specific	Specific	Standard	Standard
Algo Clusters: from deployment to production:	1 month after signing NDA+NCA	2 weeks after signing NDA+NCA	1 month after signing NDA+NCA	2 weeks after signing NDA+NCA	48 hours after signing NDA+NCA
Distribution: per World Time Zones	Colocated	Remote	Colocated	Remote	Remote
Expected Volume [signals/day]	[restricted]	[restricted]	[restricted]	[restricted]	40
Virtual Machines needed	[restricted]	[restricted]	[restricted]	[restricted]	6

Proprietary Index Example:
Autoscalp Multi-Benchmark Index.

Specs:
Objectives: pure alpha seeker.
Raw Algo output (raw signals) to be transformed by counterpart in trading signals.
Risk brackets are in line with other Autoscalp algos: +2%,-4%.

Brief:
Composed of 40+ benckmark indexes, only tradable ones.
Iso-risk weighted, same weights on all components.

Dynamics:
Basket redesigns itself every 48 hours by a pool of autonomous proprietary algos.

Failsafe and Index Protection:
Index is self protecting against typical indexes flaws.
[ I.E. SP&500 flaw: 1% of its components move S&P500  90% of times ]
Many other forms of index flaws exists:
some are known some unknown some are even knowingly
unacknowledged by exchanges thus a specific proprietary module
takes care of fully exploiting this highly arbitrable grey area.
(Failsafe Sniffer Algo).

Index Neutrality:
The index refreshing process is completely automated and uneditable.
Every 48 hours all Multi-Benchmark Index components get replaced.
During the index refreshing process, process modifications are locked out
to avoid attempts to bias the index toward any direction by any possible stake holder.

Why neutrality and lockout is needed:
There is a whole sector of Bank Pseudo-advisors, Fund pseudo-advisors , even pseudo-robot advisors,
faking objective bias while are pushing their ETFs insetead.
Then the Locking-out of the modifications during Index rebalancing was a needed feature.

The Confusion Matrix monitoring our Machine Learning Process :

Confusion Matrix measures the performance of a model.
As samples belong to two classes: forecasts can be correct (if profits) or wrong (if loses).
As the classifier predicts for a given input sample,
computing the Confusion Matrix of 2.8+ million algorithms would be totally useless.

So what we do is keep the Confusion Matrix up to date given the algorithms running at a given time.
Apparently other variables to consider would kick in such as:  size and money management etc.
but these are just applied techniques and methods to use the algos, not the algo themselves.

That assumed, the expected instantaneous Confusion matrix of the Algos in use are (rounded):
64 True Positives : The cases in which we predicted YES and the actual output was also YES.
81 True Negatives : The cases in which we predicted NO and the actual output was NO.
36 False Positives : The cases in which we predicted YES and the actual output was NO.
19 False Negatives : The cases in which we predicted NO and the actual output was YES.

which means:
64 correct signal: result was profit.
81 correct don’t trade signal: would have lost money if traded ignoring the don’t trade filter.
36 bad signal: result was a loss.
19 bad don’t trade signal: would have profited if traded ignoring the don’t trade filter.

Accuracy for the Matrix can be computed  by taking average of the values lying across
the “main diagonal”. Although We feel is is more easily understandable as posted above.

Note:
While algos looks over-specialized on “don’t trade” signals, that just depends from Expertise:
Forecasting the perfect moment and price to enter a deal is difficult
while expecting when is wise not to trade is easier:
I.E. don’t position ahead of a News Release or during a “breaking news”…

A likely objection: 64% predictability looks “too low”.

Consider it includes the broker spread and, believe us or not, is not that low.
Our Research took 12 years to go from 44% to 56% and 5 more years to rise above 62%

Corollary: anyone asserting to be owning an algo whose reliability is higher than 64%…
must prove: he did same research effort ,  did that consistently and without overtarining.

Example: think of an algo winning an algo competition whose accuracy is 90%.
Our Expertise comes in place. The top 5% of algo competition winners is so overtrained
on that market moment that would make an absurd disaster if used in real market
in a different market condition. Usually the top 5% winner algos,in the long term,
either becomes a decent random signal generators or even contrarians,
Unless someone updates it… but in that case, isn’t same algo at all.

Classic example of Overtraining:

going long on S&P500 while US Government QE was in place,
which boosted (broke) the index to climb only since 2008 ,
just 5% of the index component kept the index rise constantly.
Then such signal is not machine learning at all, is a pure loop:
believing policy won’t change means  and face a performance disaster in case it does.

Conclusion:
We expect an Asset Managers performance disaster in the close future:
because all machines since 2008 were trained to expect S&P500 to keep rising.
Which is why We keep our algos constantly under surveillance by the Confusion Matrix instead.