15-Session Programme · 25 Minutes Per Day · Free · For leaders in Nigerian crude, petroleum-products and gas trading
WEEK 1 OF 3 · THE LEADERSHIP MANDATE, EVIDENCE AND INTELLIGENCE · SESSIONS 1–5 · WORK AT YOUR OWN PACE
Orientation · The Intelligent Trading BusinessSESSION 1
The Intelligent Trading Business and the Reader's Toolkit
Why This Matters
A trader can agree a price within minutes. The cargo behind it may need weeks of financing, documentation, transport, inspection, discharge and collection. In that interval your company carries a delivery obligation, a counterparty exposure, a funding requirement and conditions that can invalidate the original economics. Artificial intelligence can help there, but only where leadership has established the evidence, authority and operating discipline that make the help useful. Over fifteen sessions you will build a signed 90-Day AI Trading Plan for your own company around one live decision you choose today, and the latest time at which better information could still change it. That deadline is where your programme begins.
25-Minute Module
0:00 – 5:00
The argument, and the finish line
- This programme is for the executive who carries the commitment and remains accountable for the result.
- THE DOCTRINE
An intelligent trading business can explain every material commitment before it makes it, every exposure while it carries it and every receipt when it closes it.
- The three weeks: 1 The Leadership Mandate, Evidence and Intelligence; 2 Crude, Products, Gas and Physical Execution; 3 Risk, Treasury, Trust and the Leadership Charge. The finish line: a signed 90-Day AI Trading Plan, fourteen sections written one per session, signed in Session 15.
- House rules: write about your own company; bring a real decision; no value claim without a baseline; no decision without a named owner; report a blocker the same day. The pass mark is 70.
- Said once: this is leadership training, not investment, trading, legal, tax or engineering advice. The frameworks, models and decision records are my proposed operating designs, not official standards or validated benchmarks. Worked examples use assumed inputs, not company results, market prices or promises of profitability. Contract, quality, credit, hedge, funding, sanctions and regulatory decisions stay with the people authorised to make them.
Bring the decision that is waiting, not the technology you have been offered.
5:00 – 9:00
One executive problem, different exposures
- The book combines crude oil, refined petroleum products and gas because they share an executive problem: the commercial promise must survive physical execution and become collected value.
- They differ in exposure. Crude carries grade and pricing-window exposure; products carry stock, specification, route and customer-payment exposure; gas carries delivery, energy-content, nomination and flexibility exposure.
- THE DOCTRINE
A common architecture is useful; a common model that ignores those differences is dangerous.
- The Petroleum Industry Act, 2021 (sections 4 and 29) establishes distinct upstream and midstream and downstream regulatory institutions. Its institutional provisions are a starting point for assessing the requirements of your activity, product, facility and transaction, not a universal trading licence. The NMDPRA's public permit portal includes application, arrival-notification, product-certification and permit-validation functions, which reinforces the need to connect each trade with its actual regulatory evidence.
- Where an example comes from an activity you do not run, apply it to the nearest equivalent decision in your business.
Name the exposure your business actually carries before you borrow anyone else's model.
9:00 – 13:00
Four mechanisms, and the questions that discipline them
- AI can retrieve a contract passage, extract a pricing formula, estimate late-payment risk or compare delivery schedules. Those are different capabilities, and each task should go to the mechanism that can be evaluated for it.
- A language model generates text. A predictive model estimates an outcome. An optimiser searches for a feasible choice under defined objectives and constraints. A controlled calculation engine applies accepted arithmetic. A fluent paragraph is not a calculation.
- As the book puts it: I do not ask a trading executive to become a machine-learning researcher. I ask the executive to understand what the application predicts, what information it had when it predicted it, what action the business can take and what loss follows when it is wrong. Those questions prevent a demonstration from standing in for an operating result.
Ask which mechanism is doing the work before you ask how well it performs.
13:00 – 20:00
The six frameworks, adapted to trading (not external standards)
Framework In the book's terms The question it forces 1. Operating-Intelligence Loop Sense, decide, act, learn Where does our loop stop, and who owns that stage? 2. Data-Foundation Precondition A price needs benchmark, unit, currency, delivery basis, publication time and permitted use Can each number behind our decision carry its attributes? 3. Autonomy Ladder Retrieval, drafting, recommendation, narrowly authorised execution What has this task earned, and what has it not? 4. Value-versus-Risk Map Rank by materiality, reversibility, time pressure and evidence quality What does a failure cost, and can it be undone? 5. Sequencing Discipline Record, reconcile, bounded pilot, compare, then expand What evidence comes before more authority? 6. Operations-Excellence Lens Fewer material contract omissions, earlier exposure recognition, feasible deliveries, fewer investigated discrepancies, more dependable cash planning Which result will we accept as proof? - IN THE PHYSICAL VIEW · THE WORKED CASE
- Physical view. In a fictional product importer, an alert says a vessel will arrive late. It is useful only if a person can change the receiving plan, notify a customer or secure a contractual position before the deadline. If nobody can, the loop breaks at act.
IN THE COMMERCIAL, RISK AND CASH VIEWS · THE PARALLEL- Commercial, risk and cash. The same delay moves the customer delivery, the receipt and the funding need. If the late collection never returns to the margin assumption behind the next cargo, the loop breaks at learn. Counting alerts measures output; counting accepted interventions and their consequences measures whether the loop works.
- The ladder. Approval to summarise an agreement does not authorise a price commitment, hedge order, payment or cargo release. An executing application needs an action limit, approved tool set, transaction identity, review route and recovery procedure. A human click helps only when the reviewer can inspect the evidence, understand the consequence and reject the action.
- THE DOCTRINE
A high potential benefit does not offset an unresolved mandatory condition. The company cannot make an unauthorised activity acceptable by assigning it a financial penalty in the optimiser.
Find the broken stage of the loop before you price the clever one.
20:00 – 25:00
The decision record, and Section 1
- The Operations-Excellence Lens ends in the decision record, which connects a model to a decision another competent person can examine.
The decision record holds What it settles Question and evidence cut-off What was decided, and what could be known at the time Alternatives and recommendation The authorised choices and the one proposed Uncertainty What remained unresolved when the choice was made Authority and accepted action Who could decide, and what was actually done Realised outcome and review date What happened, and when the choice is examined again - Before you close the laptop, write your live decision, the latest time at which better information could change it, and the role that owns it.
- Section 1 of your plan is Our trading business, the one live decision we will carry, its latest useful time and where its loop breaks.
The decision you name today is the one you will defend in Session 15.
Key Terms
Latest useful decision time
The latest time at which better information could still change a decision. The book makes it the starting point of any AI programme.
Operating-Intelligence Loop
Sense, decide, act and learn: valid evidence, authorised alternatives, an approved instruction, and a result reconciled with the original decision.
Autonomy Ladder
The proposed progression from read-only retrieval to drafting, recommendations and narrowly authorised execution, each step earned by evidence.
Decision record
Question, evidence cut-off, alternatives, recommendation, uncertainty, authority, accepted action, realised outcome and review date.
Self-Check — Before You Submit
1. A vendor's assistant summarises your supply agreements well in a pilot. The vendor proposes that it now also places the related hedge orders. What does the Autonomy Ladder say?
- A. Approve it, since the summaries proved it accurate
- B. Approve it if a trader clicks to confirm each order
- C. No: summarising does not authorise a hedge order; execution needs its own evidence and controls
- D. Reject AI for hedging permanently
C. No: summarising does not authorise a hedge order; execution needs its own evidence and controls — Each step on the ladder needs evidence for that specific task. A confirming click is tempting, but it helps only when the reviewer can inspect the evidence, understand the consequence and reject the action.
2. Which mechanism does the book describe as applying accepted arithmetic?
- A. A controlled calculation engine
- B. A language model
- C. A predictive model
- D. An optimiser
A. A controlled calculation engine — A language model generates text, a predictive model estimates an outcome and an optimiser searches for a feasible choice. Only the controlled calculation engine applies accepted arithmetic.
3. A late-vessel alert service produced many alerts last quarter. Which measure shows whether the Operating-Intelligence Loop works?
- A. The number of alerts sent
- B. How quickly alerts reached the desk
- C. The share of vessels tracked
- D. Accepted interventions and their consequences
D. Accepted interventions and their consequences — Counting alerts measures output. An alert is useful only if someone can act before the deadline, so the loop is judged on accepted interventions and what followed.
Today’s Assignment
Name your company, its trading activity and the one live decision you will carry through this programme, with its latest useful time, its owner and where its Operating-Intelligence Loop breaks.
- Your company. Its name and trading activity: crude, refined petroleum products, gas, or a function serving the trade.
- The live decision. One decision of the kind the Introduction lists, or its nearest equivalent in your business. Describe its structure and the size of the number; do not paste contract terms, prices, counterparty names or bank details.
- The deadline and the owner. The latest time at which better information could change it, and its owner by role, not department.
- Where the loop breaks. Walk it through sense, decide, act and learn. Name the stage where it stops, why, and who owns that stage.
- The record gap and the honest part. Which of the nine decision-record fields your company cannot produce today, and what it gets wrong about this decision now. No AI in use yet is normal.
- Four hundred to nine hundred words.
YOUR 90-DAY PLAN
Section 1 — Our trading business, the one live decision we will carry, its latest useful time and where its loop breaks.
How this is marked — the rubric
| Names the company, its trading activity and one specific live decision, described by structure and size without confidential detail | 20 |
| States the latest time at which better information could change the decision and an owner by role, not a department | 25 |
| Locates the loop break at a named stage with a reason and an owner | 25 |
| Lists the decision-record fields the company cannot produce today and names a real weakness in the student's own company | 30 |
| Total · pass mark 70 | 100 |
Submitting session 1
Write your answer in Word or Google Docs. The first line must read
Write your answer in Word or Google Docs. The first line must read
Session: 1, or the submission is refused. Upload the Word file (.docx), or switch to the Google Doc tab and paste your document’s share link.Marked against the rubric above and returned as a written letter, saved to your progress page. Do not submit until you have passed the self-check above.Capabilities Covered
Decision-first thinkingThe latest useful decision timeMatching task to mechanismThe six toolkit frameworksWriting a decision record
Part One · The Trading Leadership MandateSESSION 2
Lead the Whole Trade
Why This Matters
In a trading business, no single function holds the whole trade. Risk accumulates when trading, treasury, operations, risk, legal, credit and finance each work from a different version of the same trade, and the company carries a risk no single team recognises. When the chief executive hears about a disagreement only after the commitment, the remaining choice may be how to absorb it. This session is about the mandate that defines what your company may commit to, the disagreements that must reach you before approval, the incentives that shape what your people record, and the first piece of AI work I recommend: a daily exception brief that keeps the trade, the cargo, the exposure and the cash in one view.
25-Minute Module
0:00 – 5:00
The commitment belongs to the business
- THE DOCTRINE
A trader may initiate a purchase, but the company makes the commitment.
Function Its part in the commitment Trading Initiates the purchase Treasury Funds it Operations Moves it Risk Measures it Legal Defines the obligations Credit Determines acceptable customer exposure Finance Reconciles the economics - If those functions work from different versions of the trade, the company can carry risks no single team recognises.
- Applied to your decision from Session 1: which of these seven functions touch it, and which of them would describe it differently today?
If seven functions hold seven versions of one trade, the company holds none.
5:00 – 11:00
The mandate, and the disagreements that must reach you
- Create an executive mandate for each trading strategy. State its products, geography, permitted transaction structures, funding arrangements, exposure limits and excluded activities. Then define what makes an opportunity executable.
- An approved sales ambition without funding or delivery conditions is incomplete. It tells the desk what to want, not what it may commit to.
- The chief executive should receive the material disagreements between functions. The book gives two examples, seen here from both sides.
- IN THE PHYSICAL VIEW · THE WORKED CASE
- Physical view. Operations knows that the receiving window promised to the counterparty is unavailable. The cargo can be bought; it cannot yet be received as promised.
IN THE COMMERCIAL, RISK AND CASH VIEWS · THE PARALLEL- Commercial, risk and cash. The trader's margin estimate assumes collection in thirty days while treasury funds sixty. The margin presented for approval does not carry the funding the business will actually bear.
- THE DOCTRINE
Disagreements between functions are decisions to resolve before commitment, not footnotes to add after approval.
- Neither function in these examples is wrong about its own facts. The failure is that the difference reached the approver late, or not at all.
An opportunity is not executable until funding and delivery say so.
11:00 – 15:00
Incentives shape the record
- Compensation based only on booked margin can encourage premature recognition and delayed disclosure of execution costs. What people are paid for can shape what they record, and when.
- Use your company's accepted accounting and remuneration arrangements, but make the incentive effects explicit. Evaluate collections, claims, control behaviour and the treatment of late information alongside commercial contribution.
- Do not punish an early, evidenced warning as if it were the loss itself. A person who identifies a duplicate stock commitment has protected the company. A person who repeatedly makes unsupported commitments needs a different response.
- AI can help distinguish the records; leadership must judge the conduct. A model can show who raised the warning and when. It cannot decide what the company rewards.
Thank the messenger in public, or the next warning will arrive after the loss.
15:00 – 21:00
AI's first leadership assignment: the daily exception brief
- The first AI assignment I recommend is a daily exception brief connecting the trade, cargo, exposure and cash records. For every exception it shows five things.
The brief shows The question it answers Magnitude How much is at stake? Evidence What supports the exception, and from which source? Decision deadline By when must someone act? Responsible owner Who acts, by role? Authorised alternatives What may that owner do? - Divide the work by mechanism. The assistant drafts the brief from approved sources. Deterministic checks calculate limits. Owners resolve the exception in the source workflow, not in the brief.
- The brief must distinguish an absent update from a confirmed unchanged position. Design its language so that uncertainty remains visible.
- THE DOCTRINE
A silent vessel feed is not evidence that the ship has stopped. A missing customer receipt is not proof of default.
Write the brief so that silence reads as unknown, never as fine.
21:00 – 25:00
Reconstruct a trade, and Section 2
- Select a recently completed trade and reconstruct the information available at approval. Identify the first event that changed its economics and when management learned about it. Compare that time with the last useful intervention date.
- The interval between them is a more useful starting point than a general demand for predictive AI. It tells you where earlier information would have been worth having.
- At the next review, require a named owner for every unclosed material assumption: who can confirm the receiving window, the final price, the funding draw and the collection route. Shared responsibility should produce coordinated ownership, not an ownerless decision.
- Before you close the laptop, write the one disagreement between functions that most often reaches you late. Section 2 of your plan is The mandate, the disagreements between functions and the interval between the first change and the last useful intervention.
Measure the gap between the first change and the last chance before you buy a forecast.
Key Terms
Executive mandate
The statement for a trading strategy of its products, geography, permitted transaction structures, funding arrangements, exposure limits, excluded activities and what makes an opportunity executable.
Material disagreement
A difference between functions about a trade, such as the collection period or the receiving window, that must be resolved before commitment rather than noted after approval.
Daily exception brief
A brief connecting trade, cargo, exposure and cash records that shows, for every exception, magnitude, evidence, decision deadline, responsible owner and authorised alternatives.
Last useful intervention date
The latest date at which management could still have acted to change a trade's economics, compared with when it learned of the first change.
Self-Check — Before You Submit
1. A trader's margin estimate for a fictional cargo assumes collection in thirty days. Treasury will fund sixty. The deal is due for approval this afternoon. What should happen?
- A. Approve it and let finance adjust the margin afterwards
- B. Resolve the difference before commitment, with the evidence from both functions
- C. Approve it, since the trader owns the margin estimate
- D. Defer the trade until a predictive model can estimate collection
B. Resolve the difference before commitment, with the evidence from both functions — The book treats such differences as decisions to resolve before commitment, not footnotes after approval. Waiting for a predictive model is tempting but unnecessary: the disagreement is already known, and it needs a decision, not a forecast.
2. The vessel-tracking feed for a cargo has sent no update for a day. How should the daily exception brief describe the position?
- A. As an absent update, with the uncertainty visible
- B. As unchanged, since nothing has been reported
- C. As a stopped vessel requiring immediate escalation
- D. It should leave the cargo out until the feed returns
A. As an absent update, with the uncertainty visible — A silent vessel feed is not evidence that the ship has stopped, and it is not a confirmed unchanged position either. The brief must distinguish an absent update from a confirmed one so the owner sees the uncertainty.
3. Under the book's design for the daily exception brief, who calculates the exposure limits?
- A. The language assistant that drafts the brief
- B. The responsible owner, by hand
- C. The counterparty
- D. Deterministic checks
D. Deterministic checks — The assistant drafts the brief from approved sources, deterministic checks calculate limits and owners resolve exceptions in the source workflow. Letting the drafting assistant calculate limits assigns arithmetic to a text generator.
Today’s Assignment
Write the executive mandate for the trading strategy your decision belongs to, name the disagreements that reach your chief executive late, reconstruct one completed trade and assign an owner to every unclosed assumption behind your decision.
- The mandate. For the strategy behind your decision: products, geography, permitted transaction structures, funding arrangements, exposure limits, excluded activities and what makes an opportunity executable in your company.
- The disagreements. The material disagreements between functions that currently reach your chief executive late or not at all. Say which functions, and what the disagreement usually concerns.
- One completed trade. Reconstruct a recently completed trade: the first event that changed its economics, when management learned of it and the last useful intervention date. Describe its structure and the size of the number, not its confidential terms.
- The owners. Every unclosed material assumption behind your Session 1 decision, each with an owner by role.
- The honest part. Where your business's incentives or habits delay disclosure today.
- Four hundred to nine hundred words.
YOUR 90-DAY PLAN
Section 2 — The mandate, the disagreements between functions and the interval between the first change and the last useful intervention.
How this is marked — the rubric
| Writes a complete mandate covering all seven elements, tied to the Session 1 decision | 25 |
| Names specific material disagreements between functions and how late they reach the chief executive | 20 |
| Reconstructs a completed trade with the first change, when management learned of it and the last useful intervention date | 30 |
| Assigns an owner by role to every unclosed material assumption and names a real weakness in the student's own company | 25 |
| Total · pass mark 70 | 100 |
Submitting session 2
Write your answer in Word or Google Docs. The first line must read
Write your answer in Word or Google Docs. The first line must read
Session: 2, or the submission is refused. Upload the Word file (.docx), or switch to the Google Doc tab and paste your document’s share link.Marked against the rubric above and returned as a written letter, saved to your progress page. Do not submit until you have passed the self-check above.Capabilities Covered
Writing a trading mandateSurfacing disagreements before commitmentReading incentive effectsDesigning a daily exception briefReconstructing a completed trade
Part One · The Trading Leadership MandateSESSION 3
One Architecture and the Four Views of Every Trade
Why This Matters
An exporter selling crude, an importer distributing products, a refinery offtaker, a regional merchant and a gas aggregator all trade energy, but they do not carry the same economics, and a model built for one can mislead another. Before your company selects any model, it must classify what it actually does in the trade behind your decision and keep that trade's value chain visible from opportunity to closure. Then comes the book's operating model: four linked views of every trade. When your commercial, physical, risk and cash records disagree, the question is not which one is right but whether the business can explain the difference. Today you learn to require that explanation.
25-Minute Module
0:00 – 5:00
Classify the activity before selecting the model
- The book's archetypes: an exporter selling crude, an importer distributing products, a refinery offtaker, a regional merchant and a gas aggregator. They have different economics.
- Public industry descriptions support the existence of integrated physical trading, storage, transport and distribution activities. They do not establish any unnamed company's profitability, controls or use of AI, and their advertised volumes are not industry benchmarks.
- For every strategy, record whether the business takes title, acts as agent, holds inventory, charters transport or supplies customers under continuing obligations.
- Describe the optionality actually permitted. A right to choose a delivery date differs from an obligation to deliver on a date chosen by the buyer. An agency fee differs from a merchant margin exposed to price and credit.
Know whether you own the barrel before you model its price.
5:00 – 10:00
Keep the value chain visible, with one trade identity
- Map the stages: opportunity, approval, confirmation, financing, nomination, movement, inspection, invoicing, settlement and closure. Some overlap; their records should still be distinct, so the company can identify the accepted state of a transaction after an interruption.
- Use one trade identity with separate linked records for contracts, legs, physical lots, hedges, invoices and cash events. A cargo split between customers keeps its parent identity. A hedge covering several cargos has an allocation basis. An internal transfer should not create artificial external revenue or duplicate exposure.
Activity Attributes the record must carry Crude Grade, accepted assay, quantity tolerance, benchmark, differential and pricing window Products Specification, batch status, saleable quantity and the actual delivered basis Gas Composition or accepted energy basis, nomination period, delivery point, capacity rights and imbalance or flexibility terms - Keep common fields consistent without erasing product-specific evidence. A generic quantity field is insufficient when one trade is priced per barrel, another per tonne and another per energy unit. Conversion requires accepted inputs and a stated basis.
One trade, one identity, many linked records.
10:00 – 16:00
The four views of every trade
- The book proposes four linked views. Their reconciliation is the trading operating model.
View Records How its number may legitimately differ Commercial Obligations and price formulas Quantity may be a tolerance range before loading Physical Product, quality, location, quantity and movement Quantity may be provisional pending inspection Risk Sensitivities, limits, credit and contingencies May use a controlled estimate with its uncertainty Cash Payments, receipts, collateral, restrictions and funding May follow a contractual provisional invoice - THE DOCTRINE
The views do not have to contain identical numbers at every moment. They must explain differences, and each number needs its own status: confirmed, provisional, estimated or disputed.
- IN THE PHYSICAL VIEW · THE WORKED CASE
- Physical view. The book's fictional company buys 100,000 barrels. The logistics record holds a vessel parcel and a depot receipt for the same product at different stages. A careless join counts 200,000 barrels of inventory. The correct event model records movement between states; it does not sum every observation of a cargo as a separate asset.
IN THE COMMERCIAL, RISK AND CASH VIEWS · THE PARALLEL- Commercial, risk and cash. The purchase commitment and the received inventory describe the same exposure before and after receipt. The risk engine needs a defined transition so it neither drops the exposure early nor carries it twice. The daily reconciliation shows those transitions, including incomplete or disputed ones.
Ask what explains the difference, not which number wins.
16:00 – 21:00
The discrepancy workflow, the board's evidence and the first pilot
- For each discrepancy, record the competing values, their sources, the accepted basis, the affected decisions, materiality and the owner. Preserve the original records when correcting an interface, and identify whether the problem was timing, identity, unit, interpretation or a true unexplained difference.
- THE DOCTRINE
AI can group recurring causes and rank unresolved cases. It should not overwrite a custody quantity or approve an accounting adjustment. The competent owners decide treatment.
- Their accepted resolution becomes labelled evidence for later evaluation, which is how today's reconciliation becomes tomorrow's evaluation evidence.
- For the board, report unresolved exposure with age and consequence, showing what is confirmed, provisional, estimated and disputed. Do not hide a long-standing uncertain position inside a net total that appears stable. Include one material disagreement resolved through evidence; the strongest improvement may be a commitment declined because its basis could not be established in time.
- For the first pilot, choose one business line and one recurring contract structure, and build its event and evidence map before any predictive score. Measure identity errors, unexplained state changes, duplicate commitments and time to accepted reconciliation. A dashboard that reproduces conflicting spreadsheets more quickly has not solved the architecture problem.
A declined commitment can be the best evidence that the architecture works.
21:00 – 25:00
Write it down, Section 3, and Pop Quiz 1
- Before you close the laptop, write the four views of your decision on one page, and mark each number confirmed, provisional, estimated or disputed.
- Section 3 of your plan is Our activity classified, the value chain mapped and the four views of our decision reconciled.
- Pop Quiz 1 opens now: five questions, closed book, on Sessions 1 to 3. Sit it before you start Session 4.
A number without its status is an argument waiting to happen.
Key Terms
Four views of a trade
The book's proposed operating model: commercial, physical, risk and cash records of the same trade, linked and reconciled so that differences are explained.
Trade identity
One stable identifier for a transaction, with separate linked records for its contracts, legs, physical lots, hedges, invoices and cash events.
Number status
Whether a figure is confirmed, provisional, estimated or disputed, carried with the number so that no estimate is read as a fact.
Discrepancy record
Competing values, sources, accepted basis, affected decisions, materiality, owner and cause type: timing, identity, unit, interpretation or true unexplained difference.
Self-Check — Before You Submit
1. A fictional company's inventory report shows 200,000 barrels after it bought one 100,000-barrel cargo. The logistics record holds a vessel parcel and a depot receipt for the same product. What is the most likely cause?
- A. A careless join that summed two observations of one cargo as separate assets
- B. A measurement loss that must be written off
- C. A second, unrecorded purchase
- D. Fraud at the depot
A. A careless join that summed two observations of one cargo as separate assets — The correct event model records movement between states rather than summing each observation. Treating the difference as fraud or a write-off is the tempting error: a discrepancy is a question to resolve with evidence, not a conclusion.
2. An AI service has grouped recurring discrepancies between depot receipts and custody records. What may it do next?
- A. Overwrite the custody quantity with the depot figure
- B. Approve the accounting adjustment
- C. Rank unresolved cases for the competent owners to decide
- D. Close discrepancies below a size threshold
C. Rank unresolved cases for the competent owners to decide — AI can group recurring causes and rank unresolved cases. It should not overwrite a custody quantity or approve an accounting adjustment; the competent owners decide treatment, and their resolution becomes labelled evidence.
3. Which is the book's measure that a first architecture pilot has earned expansion?
- A. A dashboard that reproduces existing spreadsheets more quickly
- B. The number of predictive scores produced
- C. Positive feedback from the trading desk
- D. The record has become dependable enough for downstream analysis
D. The record has become dependable enough for downstream analysis — The pilot is judged on identity errors, unexplained state changes, duplicate commitments and time to accepted reconciliation. A faster dashboard of conflicting spreadsheets is tempting to report, but it has not solved the architecture problem.
Today’s Assignment
Classify the activity behind your decision, map its value chain with one trade identity, name the attributes it needs, set out its four views with the status of each number and record one real discrepancy.
- The classification. Whether your company takes title or acts as agent, holds inventory, charters transport or supplies customers under continuing obligations in this trade, and the optionality actually permitted.
- The value chain. From opportunity to closure, with one trade identity and the linked records at each stage. Say where your business loses the accepted state after an interruption.
- The attributes. The product-specific attributes your decision needs (crude, products or gas) and which your records do not carry today.
- The four views. Commercial, physical, risk and cash for your decision, each number marked confirmed, provisional, estimated or disputed. Give structure and size, not confidential values.
- One discrepancy. A current or recent one, with its competing values, sources, cause type, affected decision, materiality and owner by role.
- Four hundred to nine hundred words.
YOUR 90-DAY PLAN
Section 3 — Our activity classified, the value chain mapped and the four views of our decision reconciled.
How this is marked — the rubric
| Classifies the activity on title or agency, inventory, transport, continuing obligations and permitted optionality | 20 |
| Maps the value chain with one trade identity and linked records, and names the product-specific attributes missing today | 25 |
| Sets out all four views of the decision with a status for every number and explains the differences | 30 |
| Records one real discrepancy with competing values, sources, cause type, affected decision, materiality and an owner by role | 25 |
| Total · pass mark 70 | 100 |
Submitting session 3
Write your answer in Word or Google Docs. The first line must read
Write your answer in Word or Google Docs. The first line must read
Session: 3, or the submission is refused. Upload the Word file (.docx), or switch to the Google Doc tab and paste your document’s share link.Marked against the rubric above and returned as a written letter, saved to your progress page. Do not submit until you have passed the self-check above.ASSESSMENT NOW OPEN
Pop Quiz 1 — The Whole Trade and Its Four Views opens now. Sit it before you start the next session — it is how you find out whether the doctrine landed, and the practice papers may be resat as often as you like.
Sit Pop Quiz 1 →Closed book, no notes, and Coach Ritchie switches himself off while you sit it.Capabilities Covered
Classifying a trading activityOne trade identityThe four views of a tradeRunning a discrepancy workflowBoard reporting on uncertain exposure
Part Two · Evidence and IntelligenceSESSION 4
Build an AI-Ready Trading Record
Why This Matters
Every model your company might buy will be trained and tested on its records. If those records cannot say when a number became known, which version of a contract was accepted, or whether a stock figure was actually saleable, then any result the model reports is built on sand. This session is about the record before the model: the minimum fields your decision needs, the time each input became available, the difference between missing and zero, and the simple controls that often earn value before anything clever is justified. It also introduces the book's surprising first model: one that resolves what happened, rather than predicting what will happen.
25-Minute Module
0:00 – 5:00
Evidence has time as well as content
- A forecast evaluated with information published after the decision is not an honest test of the decision process. So the record must store when each piece of evidence existed, not only what it said.
Time In plain words Event time When the thing happened Publication time When the source made it known Ingestion time When your systems received it Approval time When the business accepted it - The book's example: a contract amendment can carry a document date earlier than its receipt or acceptance. A model should not assume the amendment was available merely because its printed date precedes the decision. Use the actual evidence of availability and acceptance.
- THE DOCTRINE
A backtest should reconstruct what the application could actually know at each decision cut-off. Preserve market-data revisions.
Ask when a fact became known before you ask whether it was right.
5:00 – 11:00
The minimum record, and features that carry meaning
- The minimum record: stable party and trade identifiers, legal entity, product, quantity and basis, currency, price formula, delivery location, dates, payment terms, title provisions, evidence versions and status, plus Session 3's product-specific fields. Link the document, extracted value, reviewed correction and final accepted value.
- For market observations, retain benchmark identity, tenor, settlement basis, timestamp, source permission and any transformation. Respect licensed use: a retrieved price displayed publicly does not establish permission to redistribute a proprietary history or use it for model training.
- Features, the inputs a model learns from, must carry meaning. A customer paying after an agreed revised due date differs from one paying late against the unchanged contract. A berth delay due to weather differs from a delay due to missing documents. Aggregates must preserve those distinctions.
- Leakage is when a model is fed information it could not have had at the time. Do not use future collection status to predict collection status, and do not let an invoice closure field leak the outcome into a model meant to score the invoice at issuance. Feature review traces every input to its earliest availability.
A late payment is not a definition until someone writes one down.
11:00 – 15:00
Missing data is a decision condition, and simple controls come first
- Separate six conditions: absent, not applicable, withheld, inaccessible, stale and disputed. Each calls for a different response from the owner.
- A zero should mean zero, not unknown. Imputation, filling a gap with an estimated value, may be appropriate for a predictive model, but the operating screen should disclose that it occurred and identify the affected recommendation.
- THE DOCTRINE
Run simple controls first: duplicate identity, stale price, unsupported unit conversion, inconsistent currency and missing required approval. These controls often create useful value before a complex model is justified.
- If your decision fails one of these five controls today, that is the first fix in your plan, and it needs no model.
The cheapest control that catches the error is a good result, not a lesser one.
15:00 – 21:00
Model design: the trading event graph
- The book's proposal: the first AI model should often resolve what happened, rather than predict what will happen. Each trade, parcel, contract version, invoice, movement and payment becomes a distinct record, linked by accepted relationships: a movement fulfils a contract, a certificate describes a parcel, an invoice bills a delivery, a payment settles an invoice. Each link carries source, effective time, recorded time and confidence.
- Entity resolution decides whether two records describe the same business object. Begin with exact trusted identifiers and deterministic unit checks. For the remaining candidates, a classifier (the book names a logistic classifier as one candidate) estimates the likelihood of a match from reference overlap, quantity on a common basis, date separation, counterparty and vessel identifier, trained on adjudicated pairs including difficult near matches.
- Questions to ask: are there three outcomes (an accepted link within narrowly approved rules, a proposed link for review, no link), with thresholds set from false-link consequence and reviewer capacity? Are the two errors evaluated separately? A mistaken merge can erase a real exposure; a missed link can duplicate stock. Were whole trades held out in testing? Is a reviewer's click on a model's suggestion being treated as ground truth?
- Never overwrite original records when linking them. A reversible mapping lets a reviewer separate records again and recompute the four views of the trade.
- IN THE PHYSICAL VIEW · THE WORKED CASE
- Physical view. In Nigerian product distribution, a tank reading, a release certificate and a commercial stock balance describe different things. A depot tank can read high while the product is on hold.
IN THE COMMERCIAL, RISK AND CASH VIEWS · THE PARALLEL- Commercial, risk and cash. The quantity offered for sale must combine accepted stock, quality release, reservations and transfer restrictions. A model must not learn to treat a high tank reading as freely saleable stock. Feature definitions encode operating judgement before learning begins.
Resolve what happened before you pay to predict what will.
21:00 – 25:00
Three layers, the first audit, and Section 4
- Keep three layers: immutable received evidence, accepted business state and versioned model features. A later corrected certificate enters received evidence with its arrival timestamp and changes accepted state only through the relevant workflow.
- A recommendation should be reproducible from its decision timestamp, source snapshot, feature version, model version and policy version. If any is missing, the business cannot tell whether an error came from evidence, learning or authority.
- Audit twenty representative trades, including amended, delayed and disputed cases. Trace each material decision number to its source and availability time, and record gaps with owners. Build the first dataset around resolved identities and accepted definitions, not whichever historical export has the most rows.
- Section 4 of your plan is The minimum record for our decision, the availability of each input, the missing-data rules and the first audit.
A new model may improve tomorrow; it may not rewrite yesterday's evidence.
Key Terms
As-of evidence
The evidence actually available and accepted at a decision cut-off, established from event, publication, ingestion and approval times rather than printed dates.
Leakage
A model using information that was not available at the time of the decision it claims to support, such as an invoice closure field used to score the invoice at issuance.
Entity resolution
Deciding whether two records describe the same business object, with accepted, proposed and no-link outcomes and a reversible mapping.
Available for sale
Stock that combines accepted quantity, quality release, reservations and transfer restrictions, as distinct from a tank reading or a commercial balance.
Self-Check — Before You Submit
1. A contract amendment for a fictional supply agreement is dated before last month's pricing decision but was received and accepted a week after it. How should a backtest treat it?
- A. As available at the decision, because its printed date is earlier
- B. As not available at the decision, using the actual evidence of receipt and acceptance
- C. As irrelevant, since amendments never affect pricing
- D. As available only if legal confirms the date
B. As not available at the decision, using the actual evidence of receipt and acceptance — A document date can precede receipt or acceptance. The model should not assume the amendment was available because its printed date precedes the decision; the as-of boundary comes from the evidence of availability and acceptance.
2. Your team proposes training a model to predict invoice collection, scored at issuance, using all fields in the invoice table, including the closure field. What is wrong?
- A. The closure field leaks the outcome into a model meant to score at issuance
- B. Nothing, since more fields improve accuracy
- C. Invoice tables are too small to train on
- D. Collection cannot be predicted at all
A. The closure field leaks the outcome into a model meant to score at issuance — An invoice closure field leaks the outcome into a model meant to score at issuance. More fields is the tempting answer, but every input must be traced to its earliest availability.
3. An entity-resolution model wrongly merges two neighbouring parcels into one. What is the operating consequence the book warns of?
- A. Stock will be duplicated
- B. The model will run more slowly
- C. Nothing, if the overall match rate is high
- D. A real exposure can be erased
D. A real exposure can be erased — A mistaken merge can erase a real exposure, while a missed link can duplicate stock. The two errors are evaluated separately and weighted by consequence, not hidden in one match rate.
Today’s Assignment
Define the minimum record for your decision, name when each material input becomes available, set the missing-data rules, list the simple controls to run first and plan the audit of twenty representative trades.
- The minimum record. The fields your decision needs, including the product-specific ones, and which your company cannot produce reliably today.
- Availability. For each material input, its event, publication, ingestion and approval time where relevant, and the one most likely to be treated as available before it truly was.
- Missing data. How your business will separate absent, not applicable, withheld, inaccessible, stale and disputed values, and where a zero currently means unknown.
- Simple controls. Which of the five (duplicate identity, stale price, unsupported unit conversion, inconsistent currency, missing required approval) your company should run before any model, and which it fails today.
- The audit. How you will choose twenty representative trades, including amended, delayed and disputed cases; the gaps you expect to find; and an owner by role for each.
- Four hundred to nine hundred words.
YOUR 90-DAY PLAN
Section 4 — The minimum record for our decision, the availability of each input, the missing-data rules and the first audit.
How this is marked — the rubric
| Defines a minimum record specific to the decision and names fields the company cannot produce today | 20 |
| States the availability time of each material input and identifies a real risk of leakage | 25 |
| Separates the six missing-data conditions and names the simple controls to run first | 25 |
| Plans an audit of twenty representative trades including amended, delayed and disputed cases, with expected gaps and owners by role | 30 |
| Total · pass mark 70 | 100 |
Submitting session 4
Write your answer in Word or Google Docs. The first line must read
Write your answer in Word or Google Docs. The first line must read
Session: 4, or the submission is refused. Upload the Word file (.docx), or switch to the Google Doc tab and paste your document’s share link.Marked against the rubric above and returned as a written letter, saved to your progress page. Do not submit until you have passed the self-check above.Capabilities Covered
As-of evidence and leakageDefining the minimum recordMissing-data disciplineSimple controls before modelsQuestioning an entity-resolution model
Part Two · Evidence and IntelligenceSESSION 5
Market Intelligence and Forecasting
Why This Matters
A forecasting proposal is one your business should be ready to challenge. The exact next market price may be less useful to your company than a distribution of delivered margin, customer demand or receipt timing. A forecast is worth something only when it is tied to a decision that can still change, compared with a simple baseline, honest about its uncertainty and tested on what was known at the time. This session gives you the questions to ask of any forecast placed before you, and it separates two things a polished chart often blurs: a forecast, which estimates what may happen, and a board stress, which asks what would happen.
25-Minute Module
0:00 – 5:00
Choose the forecast the business can use
- State the target, horizon, unit and decision. A weekly demand forecast should say whether it predicts orders, dispatches, accepted deliveries or collected sales. Those are four different numbers.
- Targets can mislead. Observed dispatches can understate demand during stockouts. Historical sales at a discounted price can overstate demand at a future higher price. Preserve price, availability, restrictions and unfilled orders so the target represents the business question.
- Define a target narrow enough to test. Tomorrow's released depot demand in tonnes, next week's grade differential in dollars per barrel and a customer's payment date are different problems.
- For your decision, a forecast may not be the first need at all. If the record from Session 4 is not dependable, say so.
A forecast without a decision is a chart, not a tool.
5:00 – 10:00
A meaningful baseline, and the forecasting ladder
- Use persistence, a seasonal rule or a simple regression before a more complex method. The book proposes a deliberately ordered competition, each rung a candidate to be tested, not a recommendation.
Rung Candidate What the book says about it 1 Last accepted observation or a seasonal naive forecast The first comparator 2 Transparent regression or state-space model Product, location, calendar and lagged demand features 3 Gradient-boosted trees, with XGBoost as one candidate An established family; its research origin (Chen and Guestrin, 2016) does not establish accuracy on a Nigerian trading desk Where many related series and sufficient history exist Temporal Fusion Transformer Designed for multi-horizon forecasting; the 2019 paper supports the architecture, not performance in Nigerian energy trading Alongside it A simpler linear challenger Research (Zeng and colleagues, 2022) challenged the assumption that transformer complexity necessarily improves forecasting - Require local evidence of improvement after cost and maintenance, rather than buying architectural prestige. Attention weights and feature importance can support investigation, but neither proves an economic cause.
- An unconfirmed vessel arrival is a forecast input with uncertainty, not a known future fact.
Make the expensive model beat the cheap one on your data before you pay for it.
10:00 – 15:00
Evaluate the uncertainty
- A quantile forecast gives a level the outcome should fall below a stated share of the time: the 90th percentile should be exceeded about one time in ten. The book proposes the 10th, 50th and 90th percentiles over each permitted horizon for a replenishment model.
Teaching example: 90th percentile prediction of 1,000 tonnes Actual demand Loss Demand exceeds the prediction by 100 1,100 tonnes 90 tonnes Demand falls short by 100 900 tonnes 10 tonnes - The loss is uneven because a 90th-percentile forecast is meant to run high. That asymmetry reflects the quantile level; it does not by itself establish your stockout and inventory costs. Linking quantiles to the costs of shortage and excess works only under a simplified one-period inventory problem's stated assumptions.
- Inspect calibration, whether stated coverage matches what happens, alongside interval width. A nominal 90% interval that contains outcomes only 60% of the time is misleading; one that spans every plausible state may have little planning value.
- THE DOCTRINE
Conformal prediction can construct prediction sets under specified assumptions, including exchangeability in standard formulations; time-ordered market data do not automatically satisfy exchangeability. Measure realised coverage by period and segment, and withdraw a nominal coverage claim when that evidence fails.
Ask how often the interval was right, not how narrow it looks.
15:00 – 21:00
As-of replay, and scenario kept separate from forecast
- An as-of replay lists historical decision times, freezes the sources available and permitted at each (contract, market observations, stock, receivables, operational updates), runs the model and the comparator on that snapshot and then scores the later accepted outcome.
- Ask of any training protocol: was each rolling origin trained only on information then available? Was a final period kept untouched after choosing the architecture? Did the simulation buy at prices observable only after the order, or assume liquidity the business could not access?
- THE DOCTRINE
Historical profitability produced by hindsight is not evidence of a useful trading process.
- Avoid any universal sample-size promise. Thousands of daily rows may represent only a handful of independent disruptions.
- IN THE PHYSICAL VIEW · THE WORKED CASE
- Physical view. A fictional product distributor forecasts released depot demand. Shipment timing and depot access enter as uncertain inputs. Dispatches recorded during a stockout are flagged, because they understate demand.
IN THE COMMERCIAL, RISK AND CASH VIEWS · THE PARALLEL- Commercial, risk and cash. The board also asks what happens if shipment timing, FX funding, depot access and customer collection deteriorate together. That is a stress, a management challenge, and needs no estimated probability. Sampling freight, delay, demand and FX independently can erase the joint event that consumes cash. Report which cases are probabilistic evidence and which are management challenges.
Label a stress as a stress, however realistic its paths look.
21:00 – 25:00
Pilot, suspension, Section 5 and Practice Exam 1
- Begin with one planning horizon and a decision that can be changed. Compare with the existing rule in shadow use, then in a controlled advisory workflow. Record whether the owner used the output, why it was rejected and whether action remained feasible.
- Suspend or reduce authority when data becomes stale, material inputs change definition, interval behaviour deteriorates or errors concentrate in high-consequence cases. A language assistant may prepare a sourced market digest, but must not turn an unverified rumour into a trade instruction.
- Section 5 of your plan is The forecast our decision can use, its baseline, its uncertainty test, the stresses kept separate and the suspension conditions. With it, Week 1 of your plan is complete.
- Practice Exam 1 opens now: fifteen questions, closed book, on Sessions 1 to 5, closing Week 1. Sit it before you start Session 6.
Write the suspension rule before the pilot starts, not after it disappoints.
Key Terms
Quantile forecast
A forecast of the level the outcome should fall below a stated share of the time, such as the 90th percentile, judged by a loss that penalises misses unevenly according to that level.
Calibration
Whether a model's stated coverage matches reality: a 90% interval should contain about 90% of outcomes, measured by period and segment.
As-of replay
Re-running a model and its comparator at past decision times using only the sources then available and permitted, then scoring the later accepted outcome.
Stress
A board scenario asking what would happen under selected assumptions, labelled as a management challenge rather than given a probability.
Self-Check — Before You Submit
1. A vendor presents a forecasting model that would have made large profits over the past two years on your product book. What should you ask first?
- A. Whether the model uses a transformer architecture
- B. Whether the profits could be doubled with more data
- C. Whether an as-of replay used only information available at each decision time, with execution, costs, prices and funding the business could actually access
- D. Whether competitors already use it
C. Whether an as-of replay used only information available at each decision time, with execution, costs, prices and funding the business could actually access — Historical profitability produced by hindsight is not evidence of a useful trading process. The architecture question is tempting, but prestige is not evidence; the replay must respect the information boundary and practical execution.
2. In the book's teaching example, a 90th-percentile prediction is 1,000 tonnes. Actual demand is 1,100 tonnes. What is the loss?
- A. 10 tonnes
- B. 100 tonnes
- C. 900 tonnes
- D. 90 tonnes
D. 90 tonnes — Demand exceeded a 90th-percentile prediction by 100 tonnes, giving a loss of 90; demand of 900 against the same prediction gives 10. The asymmetry reflects the quantile level and does not by itself establish the business's stockout costs.
3. Your risk team presents a scenario in which FX funding, shipment timing, depot access and collection deteriorate together, with a small probability generated from simulated paths. How should the board treat it?
- A. As a stress and management challenge, without relying on the generated probability
- B. As a validated probability for capital planning
- C. As irrelevant because its generated probability is small
- D. As a forecast to replace the demand model
A. As a stress and management challenge, without relying on the generated probability — A scenario pack does not acquire probabilities merely because a model generates realistic-looking paths. Dismissing it because the number is small repeats the same error; the stress stands as a management challenge and should be labelled as one.
Today’s Assignment
Name the uncertain input your decision depends on, or state plainly that a forecast is not your first need, and define its target, baseline, uncertainty test, replay, separate stresses, pilot and suspension conditions.
- The uncertain input. Demand, a differential, arrival or receipt timing, or another input your decision depends on. If a forecast is not your company's first need, say so and say what is.
- The definition. Target, horizon, unit and the decision it serves, and how the target could mislead (stockouts, discounted history, definitions that changed).
- Baseline and uncertainty. The meaningful baseline it must beat, how its uncertainty would be judged, and what an as-of replay at your company would need to freeze.
- Stresses kept separate. The board stresses your business should table alongside it, labelled as management challenges, not probabilities.
- Pilot and suspension. The shadow and advisory pilot, what you would record about the owner's use of it, and the conditions that would suspend or reduce its authority.
- Four hundred to nine hundred words.
YOUR 90-DAY PLAN
Section 5 — The forecast our decision can use, its baseline, its uncertainty test, the stresses kept separate and the suspension conditions.
How this is marked — the rubric
| Defines target, horizon, unit and decision precisely, or justifies plainly that a forecast is not the first need | 25 |
| Names a meaningful baseline and how uncertainty and as-of replay would be tested | 25 |
| Separates board stresses from forecasts and labels them as management challenges | 20 |
| Specifies a shadow and advisory pilot with concrete suspension or authority-reduction conditions | 30 |
| Total · pass mark 70 | 100 |
Submitting session 5
Write your answer in Word or Google Docs. The first line must read
Write your answer in Word or Google Docs. The first line must read
Session: 5, or the submission is refused. Upload the Word file (.docx), or switch to the Google Doc tab and paste your document’s share link.Marked against the rubric above and returned as a written letter, saved to your progress page. Do not submit until you have passed the self-check above.ASSESSMENT NOW OPEN
Practice Exam 1 — The Leadership Mandate, Evidence and Intelligence opens now. Sit it before you start the next session — it is how you find out whether the doctrine landed, and the practice papers may be resat as often as you like.
Sit Practice Exam 1 →Closed book, no notes, and Coach Ritchie switches himself off while you sit it.Capabilities Covered
Defining a usable forecastInsisting on a simple baselineReading calibration and intervalsQuestioning a backtestSeparating stress from forecast