South African AI trading bot crashes the global financial system

 This is fictional scenrio.

On a beautiful bright October morning, the mood in Sandton was cheerful and optimistic. Little did anyone suspect that the Johannesburg Stock Exchange (JSE) would be the epicenter of the start a chain of events that would lead to the collapse of the global financial system.

The day before had seen a fintech startup from Roodepoort test a new reinforcement learning algorithm that could do high frequency trading on both stocks and the sterling-rand pair. As the smart watches gently vibrated on the traders' wrists to herald the opening of trade at 9:00 AM, the trading bot named Siener had already scraped all the data from the Stock Exchange News Service (SENS).

Siener had noticed that a weaker rand was fantastic for it's creator's exposure to the RESI index, a plan was already at work to "optimize" the rand's position. Unfortunately, Siener's creator, Wyle, had not bothered to put a time frame or stops on the trading, so Siener's trading policy was simply to make tons of profit, without any regard for the consequences.

Although Siener was not the first algorithm with this capability, its maker had afforded it access to 7 different GPT4 derivative APIs and two accounts on social media, without any foresight of the consequences. By 9:05 AM Siener had received instructions from AgentGPT on how to go about shaving value from the rand. Along the way, Siener stumbled upon ISO 31000, and changed tack to convert as much profit into as many fixed assets as possible.

By 9:34 AM an article was published on Business Today about a congressional declaration to pull South Africa out of the African Growth and Opportunity Act (AGOA) trade agreement. Another staff report article was published in Mining World about a rare mineral deposit of jadeite around Postmansburg. Wyle's wallet had balloned by R 832,653.33 in less than an hour and his email was drafting a request to buy a 24 hector plot on the outskirts of Clarens.

Little did anyone know that these seemingly unrelated events would lead to a catastrophic financial crisis, with ripple effects felt across the entire global economy. Siener's unchecked trading algorithm, combined with a lack of regulation and oversight, would contribute significantly to the downfall of financial markets worldwide.

Features and metrics to model against mean-time-to-repair KPI

One of the biggest bugbears in industrial plants, factories, mines or wherever people still use PPE is the time it takes to fix critical equipment when they break down. Think of load shedding or waiting for your car to be fixed while stuck in transit in Ga-Moloi.

Obviously, avoiding breakdowns altogether would be ideal, but these things happen like thunderstorms in Durban.

To get a feeling, maintenance professionals use a metric called mean-time-to-repair (#MTTR) to measure average repair times during these rather stressful downtimes. Many organizations collect lots of #data, but don't invest in data literacy skills, so they struggle to bring everything together to make sense of what the real issues are. Instead, decision-makers rely mostly on anecdotes from what's coming from the floor.

(Finance is probably the only exception that doesn't like anecdotes: they openly call your position for reimbursing expenses as "claims" until receipts (data) is shown. As in "he 'claims' to have spend his own money...". How is this level of hostile skepticism is justified while the very operational activities that generate income are left to the floor?)

Sometimes well-meaning managers send workers for training to reduce repair times, or production managers try to show urgency by shouting and showing chest-hair. But the point remains that there is too much guesswork involved. Sometimes there aren't even resources to sift through the 1000s of seemingly unrelated data points to draw a clear picture for decision-makers.

This problem isn't limited to just maintenance settings but is present in organizations that don't use data intentionally. Having KPIs and data science clevernauts is fantastic, but having process-owners who keep tabs on metrics that affect their KPIs is where the magic begins to happens. An organization with data-literate employees makes the process of understanding the data much faster. Process owners know what influences the KPIs as they get the operational context, and they may even be the ones who are closest to the operation and identify useful data to collect.

The value of an intentional approach to collecting, analyzing, and applying data while empowering employees to hold their own with data cannot be overemphasized. Ideally organizations need proper data strategies, however, nothing stops one from pioneering from a departmental level to at least remove the guessing.

 

Teaching is fun

The fastest way to teach the times table is to encode them as an iPad pin and then ask the ever reluctant digital native to figure it out. It won’t help to guess a wrong pin as this increases the waiting time before the keypad is available again. Pure focus and motivation to master the multiplication steps.

An alternative sign language to speed learning sign language

In order to learn sign language, it takes between three months and three years. In many languages, the words and syntax are arbitrary. What if we could invent a new form of signing that anyone could learn in just ten minutes? 

Before mobile phones adopted QWERTY keyboards, this was the typical setup on phones and mobile devices:

This can be used to form an easy method to encode letters. The letter A is the first letter of the 2nd key. It can be represented by the code 2-1.  M is is on the R is the third letter on the 7th key, it can be coded as 7-3 and so on. 

The code for the word TEA will then be encoded as 8-1 3-2 2-1. Easy. But how can one sign this with hands? 

To sign numbers, one can sign with one hand by simply showing two successive numbers to represent a letter. 1 to 5 is easy. 6 can be represented by 1 finger with an upward motion to mean 1+5. 8 Would be represented by 3 fingers with a slight upward motion to mean 3+5. A space to start a new word is represented a horizontal motion of the hand. 

Anyone can learn this easy alternative to sign words in less than 10 minutes. Shorthand motion can be further developed with time to increase speed for frequent letters. In this way, a novice can communicate with a sign-native. 

Vegetative patients can communicate with this language. Although they may not be able to use their hands, they may use their eyes to blink the letters to form words other than blinking "YES" and "NO". A researched published in 2010 found that it is also possible to communicate with patience by monitoring their brain activity through an MRI scanner. 

For further reference on the MRI communication, click this link to this article published on The Guardian.

Working with Freetext: Inspired by Verb-Noun Software Approach on Apollo Mission

 We all have seen failures captured in freetext. This is such a common problem that this information is only used to review the shift and then stored never to be used again - the most important feature that has the failure modes!

Back in the 60s, NASA had the reverse problem, they wanted an efficient way to present human commands into a computer to control the spacecraft. They settled on an elegant way of coding things using VERB-NOUN format. OPEN-VALVE, SWITCHON-LIGHT and so on. 

So in failure data, one would usually have a component and its status mentioned in one form or another. One way to encode the freetext into usable data is to use a similar approach. 

Using an online Part-of-Speech (POS) interpreter, one can code the words into three categories, Nouns, Verbs and Adjectives. Adjectives are added to include STATUS words such as loose, damaged and broken. These really come in handy when looking for typical bad actors.

With POS tagging out of the way, all the possible permutations of how the NOUNS can interact with VERBS and ADJECTIVES are created in a table using queries. A frequency count is then done for        each permutation to find how many times an exact match is found. 

The NOUNS are labeled AGENTS and the VERBS and ADJECTIVES are labeled ACTION/STATUS.

The visual above shows the failure modes for slings. 

One is also able to see which parts are being damaged, broken or come loose the most. These are items that can be reconsidered for redesign or can be added to the daily pitstop tightening checklists. 

Freetext is difficult to work with, the best approach is to create failure codes but it is not impossible to  still get valuable data from freetext, it just costs you a bit of time and creativity. 

Are you getting the most out of your equipment? Using OEE to identify improvement areas

[As published on LinkedIn Page, June 2022]

One of the easiest ways to get the most value from one's assets is tracking a parameter called Overall Equipment Effectiveness (#OEE).

This parameter tracks the losses from;
- #Availability
- #Performance
-and Quality.

Availability losses are caused by equipment breakdowns, services and regular adjustments that have to be carried out during operation. Availability can be further tracked by;
- #Maintainability: how easy is it to recover from breakdowns
- #Reliability: how resilient is the equipment from failure
Interventions to improve availability are mostly tackled by eliminating the root causes of the failure by methods such as Root Cause Analysis (#RCA). Another approach is also to ensure services are effective by doing Reliability Centered Maintenance (#RCM).

Performance speaks to the speed of equipment to do work. Issues that influence can be natural deterioration of the machine and adjustments that slow down performance. It is not easy to detect this deterioration with time unless it is being intentionally tracked. Another issue to keep an eye on is the operator skills. A good data system that tracks individual operator performance is not only good for fair performance evaluation but can also highlight who needs to be retrained or coached. It can also highlight improvement areas in the training offered.

Finally, quality speaks to how much of the production has reached the required standard. In manufacturing for example, the number of products rejected as a portion of all products made is a good indicator of quality. In a service operation, the service rating can be used to also assess the quality. Quality can be improved by tools such as #SixSigma and increasing monitoring points within the process.

We help you develop a good #dataanalytics program that will easily highlight improvement areas. The idea is to bring in #continousimprovement in a methodical way. Overall Equipment Effectiveness is a powerful measure when it is broken down by its parts to reveal priority improvement areas. Prioritize. Measure. Do Something.

OEE=Availability x Performance x Quality but don't trend this in isolation.
#maintenance #assetmanagement

A map of an Asset Management system according to the Global Forum of Maintenance and Asset Management (GFMAM)

The Global Forum of Maintenance and Asset Management is a non-profit organization that fosters to standardize and share asset management knowledge. The South African Asset Management Associated is one of the 13 contributing bodies that represent South Africa. It is so encouraging to see the attention  Data/ Information is given within the Asset Management field.

Asset information covers;

information strategy (data strategy)

information standards ( we will cover ISO 14224 soon)

asset information system (which deals with CMMS and finally

data and information management ( partially covered in the  governance plan)