Class 5&6

S5] Session 5 · line 1

A rule is written, a model is learned from data.

[S5] Session 5 · line 2

The luckiest coin looks like skill.

[S5] Session 5 · line 3

A pattern everyone trades stops being a pattern.

[S5] Session 5 · line 4

Fluent is not the same as right.

[S5] Two words to recognise (session 5)

Alternative data · hallucination

[S5] SEC fines Delphia and Global Predictions (18 March 2024)

$400,000 in total. Their "AI" did not exist ("AI washing"). Session 3 line: "It follows rules, it does not think."

[S5] Which two links does session 5 ask about?

Advise (reads before you buy) and Report (writes after you buy). Can a machine do these two jobs? Should it?

[S5] Apple Card (2019): who made the credit decisions?

Goldman Sachs, the US investment bank. "Created by Apple, not a bank", but every credit decision was Goldman's.

[S5] Apple Card controversy (from 7 Nov 2019)

Same tax return, but the husband's limit was 20 times his wife's. Customer service couldn't explain why; an Apple co-founder posted the same story; New York's regulator (NYDFS) opened an investigation.

[S5] What did the NYDFS find (March 2021)?

Nearly 400,000 New York applications checked: no evidence of discrimination (each limit followed the credit file), but customers got no explanation. Models learned from past lending can carry its old bias.

[S5] A rule vs a model

Rule: somebody wrote it ("IF income above €1,500 AND no missed payment THEN yes"); you can read why. Model: nobody wrote it, learned from past borrowers; no sentence to read, so someone must explain it.

[S5] Where machine learning already works in the chain

2 Check (matches your selfie to your ID) · 3 Advise (reads results, counts cars) · 5 Trade (calls, helps execute big orders) · 7 Report (drafts the email after a meeting). Hardest job: picking the shares.

[S5] Backtest

Testing a strategy on old prices. A machine tests thousands overnight; the best one always looks brilliant, but it is often the luckiest coin.

[S5] Overfitting

The model learns the past by heart, luck included, so it fails on new data. Happens when there are few examples.

[S5] Coin simulation: setup

1,000 coins = 1,000 managers with zero skill. 1 toss = 1 month (heads = beat the market). 120 tosses = 10 years. Then pick the best.

[S5] Coin simulation: result

The best coin beat the market 81 of 120 months (68%) in years 1–10, then only 31 of 60 months (52%) in years 11–15. Any coin: 50%. Luck, chosen after the fact.

[S5] First question to ask about a brilliant backtest

How many strategies did you try?

[S5] Harvey, Liu and Zhu (2016)

316 published patterns said to predict returns (counted in 2012). Many fail a test that counts the tries.

[S5] Best US share funds of 2021: how many stayed on top to 2025?

Almost none: 0% of 2021's top-quarter funds stayed in the top quarter every year to 2025 (S&P Dow Jones Indices, Persistence Scorecard, 7 May 2026). Line 2.

[S5] How a model is trained (the cat)

Examples with the answer (millions) → training (guess, check, adjust, millions of times) → test on new photos. Right: it learned. Wrong: it memorised.

[S5] Why markets are not cats: data

ImageNet: ~14 million labelled photos. One US share: at most 1,200 months of prices since 1926.

[S5] Four reasons a market is harder than a cat

1 Little data · 2 A weak signal in a lot of noise · 3 The world changes · 4 The market learns, and patterns vanish (never happens to a cat).

[S5] McLean and Pontiff (2016): 97 published patterns

Return index: in the study 100 → after the study, not yet published 74 (26% lower) → after publication 42 (58% lower).

[S5] How a pattern dies, in four steps

1 Found · 2 Published (anyone can test it) · 3 Traded (funds buy, price rises now) · 4 Copied (everyone buys earlier). The price knows; nothing left to earn.

[S5] After publication, what happens to a pattern's shares?

More trading, more short selling of the shares it says will do worse, −32% of the return (investors trading on the paper). "Financial markets 'learn' as investors learn what works."

[S5] Alternative data

Non-traditional data used to know sooner, e.g. satellite images of car parks (fewer cars → bad quarter → sell before results). Goldman Sachs research uses satellite images, phone foot traffic, app downloads, website visits, search trends, earnings-call text read by AI. The machine counts; people decide what it means.

[S5] Morgan Stanley's AI

98% of its adviser teams use its AI assistant (26 June 2024); a second tool (Debrief) takes notes in client meetings and drafts the follow-up email. Link 7, Report.

[S5] Mata v. Avianca (2023)

A passenger sues Avianca (a serving cart hit his knee). His lawyer filed six court decisions found by ChatGPT: 0 of 6 were real. $5,000 fine (22 June 2023). He filed them and asked the machine if they were real.

[S5] What should the lawyer have done?

Find each case in a legal database, i.e. check at the source, not ask ChatGPT.

[S5] Hallucination

An answer that sounds certain and is made up. Not a lie (it doesn't know), not old news, not a bug the next update removes.

[S5] Exam: backtest beat the market 15 years; the fund tried 5,000 strategies

Maybe the luckiest of 5,000.

[S5] Exam: AI writes a perfect client report with three numbers. Before sending?

Check the three numbers at the source.

[S6] Session 6 · line 1

Every rule protects against one of four fears.

[S6] Session 6 · line 2

The licence follows the verb.

[S6] Session 6 · line 3

1 licence, 27 countries.

[S6] Session 6 · line 4

The rule targets the behaviour, not the technology.

[S6] Word to recognise (session 6)

Tied agent: working under another firm's licence.

[S6] SCHUFA ruling (EU Court of Justice, 7 Dec 2023)

SCHUFA, a German credit agency, sells banks a repayment score. If the bank follows the score, the machine is deciding (an automated decision under GDPR art. 22): banned in principle.

[S6] The four fears

1 Dirty money (anti-money laundering) · 2 A product wrong for you (investor protection) · 3 A rigged market (market integrity) · 4 A firm that fails with your money inside (financial stability).

[S6] One real case per fear

Dirty money: N26 (BaFin 2021) · Wrong product: Robinhood, $70m (FINRA 2021) · Rigged market: JPMorgan, $920m (CFTC 2020, spoofing) · Firm fails: FTX, bankrupt 2022.

[S6] N26 and BaFin

German bank licence since 2016. June 2021: €4.25m fine for suspicious payments reported late. Autumn 2021: BaFin caps new clients at 50,000 a month and appoints a special commissioner inside. May 2024: €9.2m fine for the same failure.

[S6] What did BaFin do when N26 grew faster than its controls?

Limited its new clients until the controls worked (max 50,000/month) and put a watcher inside. Not another fine, not removing the licence.

[S6] EU ban on payment for order flow

MiFIR art. 39a: no broker may be paid for sending your order. Whole EU from July 2026 (exemptions until 30 June 2026). One rule, two fears: wrong product (worse price for you) and rigged market.

[S6] Trade Republic's two verbs (2018)

Buy your shares for you, and keep your shares safe. Two verbs, two licences; no licence, no first client. Own BaFin licence from 19 December 2018.

[S6] Own licence vs tied agent

Own licence: own capital, up to 6 months for an answer, keep all the margin. Tied agent: use one partner's licence; the partner answers for you fully; you share the margin.

[S6] Trade Republic 2023: take cash and lend. Which licence?

A bank licence. It became a full bank in December 2023. New verb, new licence.

[S6] Minimum initial capital by verb

Advise, pass orders €75,000 · + hold clients' assets €150,000 · trade with own money €750,000 · take deposits (a bank) €5 million. Hold is the most expensive verb.

[S6] Coinbase fine (Central Bank of Ireland, 5 Nov 2025)

€21.5m: 30 million transactions not fully checked for dirty money (Apr 2021–Apr 2022). Its US sister company ran the checks, but the licensed Irish firm is to blame. Line 2: the licence follows the verb.

[S6] Where your apps get their licence

Revolut: Lithuania (Bank of Lithuania) · Trade Republic & N26: Germany (BaFin) · crypto: Luxembourg (CSSF) · Sweden (Finansinspektionen) · France (ACPR). One of six is French; all six can serve you in France.

[S6] You live in Paris; Revolut is Lithuanian, Trade Republic German. Which rules protect you?

The same EU rules for both. Rules: EU law, the same for all 27. Supervisor: only the one at home reads your file (e.g. Bank of Lithuania fined Revolut Bank €3.5m, 4 April 2025).

[S6] EU passport

One licence lets a firm serve clients in all 27 EU countries (MiFID II art. 34, CRD art. 33, PSD2 art. 28, MiCA art. 65).

[S6] A UK-only licence: how many EU countries?

0, since 1 January 2021. Brexit ended the passport; a London fintech needs a second licence inside the EU.

[S6] Fast trading robots: banned?

No. Test them, keep a kill switch; markets can pause (MiFID II art. 17, 48(5)).

[S6] AI that decides who gets a loan: banned?

No. You can ask for a human to look again and challenge it (GDPR art. 22(3)). The AI Act makes credit scoring "high-risk" (strict rules, not a ban) from 2 Dec 2027.

[S6] WhatsApp for bankers: banned?

No. Calls and messages about trades must be recorded and kept 5 years (MiFID II art. 16(7)).

[S6] Off-channel messaging fines (27 Sept 2022)

$1.8bn in one day: SEC $1.1bn (16 firms) + CFTC $710m (11 firms). WhatsApp was never banned; the missing records were.

[S6] The EU laws you will meet

MiFID II (investing, 2018) · GDPR (data, 2018) · AML (checking clients) · PSD2 (payments) · MiCA (crypto) · DORA (IT that must not fail) · AI Act (AI). Each one names a behaviour.

[S6] Exam: a robo manages your portfolio, a partner bank keeps the money. What decides its licence?

What it does: managing your portfolio (the verb), not the name, technology or your country.

[S6] Exam: Irish investment licence, wants clients in Spain next month

Its Irish licence, used as a passport.