Ethics: personal data, artificial intelligence and selling signals.

When the system touches other people

Lessons 6.7.3 and 6.7.4 deal with the personal data that come from outside: what a scraper collects from a website and what a purchased alternative dataset contains. This lesson looks inward and ahead. A systematic trader often ends up with other people's data: the mailing list of their blog, the subscribers of a signals channel, the client accounts of a tool, the conversations of an AI assistant. And with them come three ethical questions that also have a legal answer:

1. What personal data do I process and how long do I keep them? The GDPR.
2. What obligations does the AI I use with those people have? The AI Act and, if there are investment services, MiFID II.
3. What can I say about my results and what can I not do with my recommendations? MAR, the rules on investment advice and the warnings of the CNMV (Comisión Nacional del Mercado de Valores, Spain's securities regulator), plus a question of statistical honesty that no rule fully settles: how a track record is manufactured.

GDPR: the data of a trading service

Article 5 of Regulation (EU) 2016/679 (GDPR) sets out the principles: data are collected for specified purposes and not further processed in an incompatible way (purpose limitation), they must be “adequate, relevant and limited to what is necessary” (data minimization), they are kept in a form that identifies the data subject only for as long as necessary (storage limitation) and with appropriate security11Note 1 · Regulation (EU) 2016/679 (GDPR), Articles 4(5), 5, 6, 17, 22 and 25, BOE (Spanish version; quotations are unofficial translations), accessed on October 23, 2026.. Every processing operation needs a basis in Article 6: the contract to send the signals to whoever pays for them, a legal obligation for invoicing, consent for the marketing mailing list.

A measure the regulation cites explicitly, in Article 25 (data protection by design), is pseudonymization: processing data so that they “can no longer be attributed to a specific data subject without the use of additional information, provided that such additional information is kept separately” (Art. 4(5)). The European Data Protection Board (EDPB) guidelines 01/2025 recall that pseudonymized data are still personal data and cite functions such as HMACs, with a secret kept apart, as a way to do it22Note 2 · EDPB, “Guidelines 01/2025 on Pseudonymisation”, adopted on January 16, 2025 (version for public consultation), edpb.europa.eu.. The usual mistake is to use an unkeyed hash:

import hashlib
import hmac

import numpy as np
import pandas as pd

subs = pd.DataFrame({
    "email": [f"usuario{i}@example.com" for i in (3, 17, 42, 108,
                                                  512, 999)],
    "ip": [f"203.0.113.{i}" for i in (5, 17, 23, 42, 77, 91)],
    "plan": ["básico", "pro", "pro", "básico", "pro", "básico"],
})

def hash_simple(x):
    return hashlib.sha256(x.encode()).hexdigest()[:16]

# in production, the key lives in a secrets manager (lesson 4.4.5)
CLAVE = b"clave-de-ejemplo-guardada-aparte"

def seudonimo(x, clave=CLAVE):
    return hmac.new(clave, x.encode(), hashlib.sha256).hexdigest()[:16]

subs["h_simple"] = subs["email"].map(hash_simple)
subs["seud"] = subs["email"].map(seudonimo)

# someone who has the log and a list of candidate emails
candidatos = [f"usuario{i}@example.com" for i in range(1, 1001)]
dicc = {hash_simple(c): c for c in candidatos}
dicc_k = {seudonimo(c, b"clave-adivinada"): c for c in candidatos}
print("unkeyed hash, re-identified:",
      subs["h_simple"].map(dicc).notna().sum(), "of", len(subs))
print("HMAC, without the key:", subs["seud"].map(dicc_k).notna().sum())
print("HMAC, with the key:",
      subs["seud"].map({seudonimo(c): c for c in candidatos})
      .notna().sum())

unkeyed hash, re-identified: 6 of 6
HMAC, without the key: 0
HMAC, with the key: 6

An email is predictable data: with a list of candidates (a thousand here; in reality, leaks with millions), the unkeyed hash is undone entirely. The HMAC cannot be undone without the key, and with it, it can: that is why the key is stored separately from the log and with restricted access. Pseudonymized data are still personal, but the risk of a log leak drops a lot.

Minimization comes first: the log used to measure which signals are delivered and to which plan does not need the email or the IP, nor the exact time. The plans keep their Spanish labels (básico, basic, and pro):

envios = pd.DataFrame({
    "email": subs["email"].repeat(2).to_numpy(),
    "ip": subs["ip"].repeat(2).to_numpy(),
    "ts": pd.to_datetime(["2026-10-21 08:59:58.120"] * 6
                         + ["2026-10-22 09:00:01.873"] * 6),
    "senal": ["NORU compra"] * 6 + ["MURI venta"] * 6,
})
analitico = pd.DataFrame({
    "seud": envios["email"].map(seudonimo),
    "plan": envios["email"].map(subs.set_index("email")["plan"]),
    "dia": envios["ts"].dt.date,
    "senal": envios["senal"],
})
print(list(envios.columns), "->", list(analitico.columns))
analitico.groupby(["dia", "plan"]).size()

['email', 'ip', 'ts', 'senal'] -> ['seud', 'plan', 'dia', 'senal']
dia         plan
2026-10-21  básico    2
            pro       4
2026-10-22  básico    4
            pro       2
dtype: int64

The analytical log answers what the service needs to know (how many signals are delivered each day per plan) with no emails, IPs or milliseconds. The delivery log with IPs, if it is needed for security, has another purpose, other access and a short retention period.

Erasure versus the obligation to retain

Article 17 of the GDPR gives the right to obtain the erasure of data, but its paragraph 3, point (b), excludes it when processing is necessary to comply with a legal obligation¹. A trading service has several:

• Accounting and invoices. In Spain, Article 30 of the Commercial Code (Código de Comercio) requires books, correspondence, documentation and supporting records of the business to be kept “for six years, from the last entry made in the books”33Note 3 · Spanish Commercial Code (Código de Comercio), Article 30.1, BOE-A-1885-6627 (in Spanish; the quotation is an unofficial translation), consolidated text accessed on October 23, 2026..
• Client orders and transactions, if the service is an investment firm: five years (Art. 25(1) of MiFIR) and, for recordings and communications, five years or up to seven if the authority requests it (Art. 16(7) of MiFID II)44Note 4 · Regulation (EU) No 600/2014 (MiFIR), Article 25(1), and Directive 2014/65/EU (MiFID II), Article 16(7), EUR-Lex..

flowchart TB
  P[Request] --> B{Legal basis?}
  B -- consent --> S[Erase now]
  B -- contract --> F{Contract active?}
  F -- no --> S
  F -- yes --> M[Keep]
  B -- obligation --> V{Period expired?}
  V -- yes --> S
  V -- no --> K[Block and retain]

Turned into code, with injected dates and the simplified rule that the six-year period counts from the date of the document (the bases keep their Spanish keys: consentimiento, contrato, obligación):

PLAZO_ANOS = {"contabilidad (art. 30 C. Com.)": 6}

datos = pd.DataFrame([
    ("mailing list", "consentimiento", None, None),
    ("profile and preferences", "contrato", None, False),
    ("invoice 2024-0312", "obligación",
     "contabilidad (art. 30 C. Com.)", "2024-03-12"),
    ("invoice 2019-0044", "obligación",
     "contabilidad (art. 30 C. Com.)", "2019-02-04"),
], columns=["dato", "base", "norma", "fecha_o_vivo"])

def resolver(fila, hoy):
    if fila.base == "consentimiento":
        return "erase"
    if fila.base == "contrato":
        return "keep" if fila.fecha_o_vivo else "erase"
    fin = (pd.Timestamp(fila.fecha_o_vivo)
           + pd.DateOffset(years=PLAZO_ANOS[fila.norma]))
    if hoy >= fin:
        return "erase"
    return f"retain blocked until {fin.date()}"

hoy = pd.Timestamp("2026-10-23")
for f in datos.itertuples():
    print(f"{f.dato:<24} {resolver(f, hoy)}")

mailing list             erase
profile and preferences  erase
invoice 2024-0312        retain blocked until 2030-03-12
invoice 2019-0044        erase

The mailing list is erased as soon as consent is withdrawn, and the profile, because the contract has already ended. The 2024 invoice is retained, but blocked: only for the purpose of the obligation (an inspection, an audit), not for sending advertising or training a churn model. The 2019 one has already passed its period and should have been erased without waiting for any request: storage limitation does not depend on someone claiming it.

Responsible AI: the AI Act and its dates

Regulation (EU) 2024/1689 on artificial intelligence classifies systems by risk. Its application dates are in Article 113, which Regulation (EU) 2026/1744 (the “Digital Omnibus on AI”), published in the Official Journal on July 24, 2026 and in force three days later, has amended55Note 5 · Regulation (EU) 2024/1689, Articles 4, 50 and 113 and Annex III, EUR-Lex, and Regulation (EU) 2026/1744 of the European Parliament and of the Council of 8 July 2026 (OJ L, 24.7.2026), amending its Articles 4 and 113, EUR-Lex. Texts read in the Spanish version of the EU Publications Office on October 23, 2026.:

calendario = pd.DataFrame([
    ("2025-02-02", "chapters I and II: prohibitions and literacy"),
    ("2025-08-02", "general-purpose AI, governance and penalties"),
    ("2026-08-02", "general application, Art. 50 transparency"),
    ("2026-12-02", "marking (Art. 50.2) of pre-existing systems"),
    ("2027-12-02", "Annex III high-risk (Art. 6.2)"),
    ("2028-08-02", "high-risk products in Annex I (Art. 6.1)"),
], columns=["desde", "que"])
calendario["desde"] = pd.to_datetime(calendario["desde"])

def vigentes(dia):
    c = calendario.copy()
    c["estado"] = np.where(c["desde"] <= dia, "applicable",
                           (c["desde"] - dia).dt.days.astype(str)
                           + " days")
    return c

vigentes(pd.Timestamp("2026-10-23"))

       desde                                           que      estado
0 2025-02-02  chapters I and II: prohibitions and literacy  applicable
1 2025-08-02  general-purpose AI, governance and penalties  applicable
2 2026-08-02     general application, Art. 50 transparency  applicable
3 2026-12-02   marking (Art. 50.2) of pre-existing systems     40 days
4 2027-12-02                Annex III high-risk (Art. 6.2)    405 days
5 2028-08-02      high-risk products in Annex I (Art. 6.1)    649 days

For a systematic trader, three consequences:

• A proprietary trading model is not high-risk just for being one. The high-risk categories of Annex III are a closed list; in finance they include systems for assessing the creditworthiness of natural persons or their credit score and those for risk assessment and pricing of life and health insurance⁵. Buying and selling for oneself is not listed.
• Transparency may well affect the service. Since August 2, 2026, Article 50(1) requires systems that interact directly with people to be designed so that those people know they are talking to an AI, unless it is obvious. An assistant that answers the subscribers of a signals channel is the typical case. AI literacy under Article 4, as worded by the omnibus, asks providers and deployers to take measures to support it among those who operate the systems, without requiring a specific level.
• If there are clients and investment services, MiFID II prevails. ESMA said so in its statement of May 30, 2024: the firm's decisions remain the responsibility of its management body, “irrespective of whether those decisions are taken by people or AI-based tools”, and MiFID II obligations apply to the AI tools used with clients66Note 6 · ESMA, “Public Statement on the use of Artificial Intelligence (AI) in the provision of retail investment services”, ESMA35-335435667-5924, May 30, 2024, esma.europa.eu.. If, in addition, an automated decision produces significant effects on a person (rejecting a client, for example), Article 22 of the GDPR comes into play.

Selling signals: recommendation or advice

Article 20 of MAR requires whoever produces or disseminates investment recommendations to present them objectively and to disclose their interests or conflicts77Note 7 · Regulation (EU) No 596/2014 (MAR), Articles 12(2)(d) and 20, OJ L 173, 12.6.2014, copy in the BOE (Spanish version).. The CNMV puts it this way in its guide for finfluencers: “Anyone can publish investment recommendations”, but with general requirements (who produces it, date and time, facts separated from opinions, sources, interests and conflicts visible on each channel) and additional requirements for professionals and “experts”, among them the methodology, the investment horizon and risk warnings88Note 8 · CNMV, “Finfluencers: cómo actuar con responsabilidad” (guide, May 2026), sections IV to VII, cnmv.es (in Spanish; quotations are unofficial translations).. Delegated Regulation (EU) 2016/958 details them; its Article 4(1)(i) asks experts for a list of all their recommendations from the previous twelve months, with the date and price of each one99Note 9 · Commission Delegated Regulation (EU) 2016/958 of 9 March 2016, Articles 2 to 5, EUR-Lex..

Investment advice is a different matter: a service that requires authorization. Article 9 of Delegated Regulation (EU) 2017/565 defines it by the personal recommendation: one presented as suitable for that person or based on their personal circumstances, and it clarifies that “a recommendation is not a personal recommendation if it is issued exclusively to the public”1010Note 10 · Commission Delegated Regulation (EU) 2017/565, Article 9, EUR-Lex..

flowchart TB
  M[Message] --> A{Suggests investing?}
  A -- no --> I[Information]
  A -- yes --> P{Personalized?}
  P -- no --> R[MAR recommendation]
  P -- yes --> S[Advice]
  S --> Z[Requires authorization]

The CNMV is explicit about disclaimers: without authorization, no personalized recommendations should be given on what to buy, sell or hold, and publishing “This is not investment advice” “will not protect you in these cases, nor does it exempt you from complying with the applicable regulations”⁸. In its guide for followers it lists “alerts and signals (buy/sell)” among risky content and warns that the main risk is fraud: finfluencers who provide advice without authorization or recruit clients for chiringuitos financieros, the Spanish term for boiler rooms, entities that provide investment services without being authorized1111Note 11 · CNMV, “Del like a la inversión: qué debes saber de los finfluencers” (guide, April 2026), cnmv.es, and “Chiringuitos financieros” (guide), cnmv.es, both in Spanish.. Its search engine of warnings and its registers of authorized entities are the prior check it recommends1212Note 12 · CNMV, search engine of warnings about unauthorized entities, cnmv.es..

Two frequent conflicts of interest in selling signals:

How a track record is manufactured

A signals seller shows their track record. The ethical problem lies in how that track record is built, even if every number is true. To measure it, the next cell creates 2,000 channels with no skill at all: every day they give a random buy or sell signal on an asset with no trend. After a year, the channels with an annual Sharpe above 1 are advertised, and their performance over the next six months is measured:

rng = np.random.default_rng(2026)
N, DIAS, FUERA = 2000, 252, 126
r = rng.normal(0, 0.01, DIAS + FUERA)        # asset with no trend
senal = rng.choice([-1, 1], size=(N, DIAS + FUERA))   # no skill
ret = senal * r

def sharpe(x, axis=-1):
    return np.sqrt(252) * x.mean(axis) / x.std(axis, ddof=1)

sr_dentro = sharpe(ret[:, :DIAS])
sr_fuera = sharpe(ret[:, DIAS:])
anunciado = sr_dentro > 1
print(f"advertised: {anunciado.sum()} of {N}")
print(f"average Sharpe, year of the advertisement: "
      f"{sr_dentro[anunciado].mean():.2f}")
print(f"average Sharpe, six months later: "
      f"{sr_fuera[anunciado].mean():.2f}")
print(f"all channels, year of the advertisement: {sr_dentro.mean():.2f}")

advertised: 316 of 2000
average Sharpe, year of the advertisement: 1.49
average Sharpe, six months later: -0.04
all channels, year of the advertisement: 0.02

One in six channels exceeds a Sharpe of 1 without knowing anything. Their average in the year of the advertisement, 1.49, is what their pages would show; over the following six months, that of a coin toss. The figure at the top summarizes the three numbers. In reality, the bias is worse: losing channels close, change their name and start over, and the reader only sees the survivors.

Lesson 2.5.4 provides the yardstick: the maximum Sharpe to be expected among $N$ strategies with no skill, with the approximation by Bailey and López de Prado ($\gamma$ is the Euler-Mascheroni constant and $\sigma_{SR}$ the standard deviation of the estimated Sharpe ratios):

from scipy.stats import norm

g = np.euler_gamma
esperado = sr_dentro.std() * ((1 - g) * norm.ppf(1 - 1 / N)
                              + g * norm.ppf(1 - 1 / (N * np.e)))
mejor = sr_dentro.argmax()
print(f"best channel: Sharpe {sr_dentro[mejor]:.2f} "
      f"(six months later: {sr_fuera[mejor]:.2f})")
print(f"maximum expected with no skill for N={N}: {esperado:.2f}")

best channel: Sharpe 3.06 (six months later: -1.52)
maximum expected with no skill for N=2000: 3.40

The best of 2,000 random channels has a Sharpe of 3.06, below what can be expected from the best of 2,000 coins. A track record with a Sharpe of 3 says nothing unless you know how many attempts it comes from.

Choosing the window

The second trick does not require advertising only the best: choosing the period is enough. Each channel shows its “last three months” right when they are the best of the year. With 63 sessions, the estimated Sharpe has twice the noise it has with 252, and among all the possible windows there is always a good one:

panel = pd.DataFrame(ret[:, :DIAS].T)          # one column per channel
v = panel.rolling(63)
mejor_ventana = (np.sqrt(252) * v.mean() / v.std()).max()
print(f"average Sharpe over the full year: {sr_dentro.mean():.2f}")
print(f"average Sharpe of the best window: "
      f"{mejor_ventana.mean():.2f}")
print(f"channels with some window > 2: "
      f"{(mejor_ventana > 2).mean():.0%}")

average Sharpe over the full year: 0.02
average Sharpe of the best window: 3.41
channels with some window > 2: 84%

Over the full year, the average channel has a Sharpe of 0.02; with its best three-month window, 3.41 on average, and 84% of the channels with no skill at all have some window above 2. “Our last three months” is informative only if the cut-off date was fixed before seeing the results.

Deleting the losing signals

The third trick is the simplest: not publishing, or deleting afterward, the signals that go wrong. With a mediocre channel, the one at the median of the 2,000 by Sharpe, and one trade per day:

mediano = np.argsort(sr_dentro)[N // 2]
ops = ret[mediano, :DIAS]
publicadas = ops[ops > 0]          # the losing ones “get deleted”
print(f"signals: {len(ops)}, published: {len(publicadas)}")
print(f"actual hit rate: {(ops > 0).mean():.1%}; "
      f"published hit rate: {(publicadas > 0).mean():.0%}")
print(f"actual return: {ops.sum():+.1%}; "
      f"published: {publicadas.sum():+.1%}")

signals: 252, published: 124
actual hit rate: 49.2%; published hit rate: 100%
actual return: +0.3%; published: +107.7%

The median channel is right on 49.2% of the days and gains 0.3% over the year. Without its 128 losing signals, it publishes 124 with a 100% hit rate and a sum of +107.7%. A 100% hit rate is impossible in an honest track record and easy in an edited one. That is why the complete list of recommendations for twelve months that Article 4(1)(i) of Delegated Regulation 2016/958 requires of experts matters more than any summary: with all the signals, dated and with their price, the hit rate goes back to the real one.

The letter that is right ten times in a row

The last trick is old and needs no market. A prediction (“up” or “down”) is sent to 1,024 people, half receive one and half the other; the next week it is repeated only with those who received the right one:

destinatarios = np.arange(1024)
for semana in range(1, 11):
    sube = destinatarios[: len(destinatarios) // 2]
    resultado_sube = rng.random() < 0.5
    destinatarios = sube if resultado_sube else \
        destinatarios[len(destinatarios) // 2:]
    if semana in (1, 5, 10):
        print(f"week {semana:>2}: {len(destinatarios):>4} people "
              "without a single miss")

week  1:  512 people without a single miss
week  5:   32 people without a single miss
week 10:    1 people without a single miss

One person has received ten correct predictions in a row, a probability of 1 in 1,024 if they were random, and has no way of knowing there are another 1,023 letters. It is the extreme version of the advertised channels: whoever looks at the track record does not see the denominator. The question that dismantles it is always the same: how many attempts were there and where are all of them?

Exercises

1. Repeat the channel simulation advertising only those with a Sharpe above 2. How many are there, what average Sharpe do they show and what Sharpe do they have afterward?

for umbral in [1, 2]:
    a = sr_dentro > umbral
    print(f"threshold {umbral}: {a.sum():>3} channels, "
          f"in-sample {sr_dentro[a].mean():.2f}, "
          f"out-of-sample {sr_fuera[a].mean():.2f}")

threshold 1: 316 channels, in-sample 1.49, out-of-sample -0.04
threshold 2:  37 channels, in-sample 2.35, out-of-sample -0.10

2. A channel publishes 40 signals in a year; it says it got 30 right. It gives you the complete list with dates and prices and, when you check it, 12 of the 30 “winners” had an entry price that was not reached that day. Compute the real hit rate if those 12 do not count as trades and if they count as misses.

total, aciertos, imposibles = 40, 30, 12
print(f"declared: {aciertos / total:.0%}")
validas = total - imposibles
print(f"excluding impossible: {(aciertos - imposibles) / validas:.0%}")
print(f"impossible ones as misses: {(aciertos - imposibles) / total:.0%}")

declared: 75%
excluding impossible: 64%
impossible ones as misses: 45%

3. A paid Telegram channel sends the same three ideas every morning to all its subscribers. The administrator also opens a private chat for those who pay for the “premium” plan, where they ask about the subscriber's capital and profile and say how much to buy. Which part is a recommendation and which part is advice? What changes if the chat is handled by an AI assistant?