ESMA and CFDs: leverage, close-out and negative balance.

From the ESMA decision to the CNMV resolution

Lesson 1.2.4 presented the limits on CFDs for retail clients and computed how much movement closes a position in EURUSD. This lesson goes to the sources and to the detail that decides the outcome of an account: which underlying falls in which band, how the 50% close-out is measured, what happens with a price gap and where the figure in the warning comes from. The framework is the EU one as applied in Spain.

The product intervention power came with Regulation (EU) 600/2014 (MiFIR), applicable since January 3, 2018: its Article 40 allows ESMA to temporarily restrict a product throughout the EU and Article 42 allows each national authority to do so in its own country. On May 22, 2018 ESMA adopted two decisions: banning binary options for retail clients from July 2 and restricting CFDs from August 1 (Decision (EU) 2018/796), for three renewable months. According to the CNMV (Comisión Nacional del Mercado de Valores, Spain's securities market regulator), ESMA renewed them three times. To give them stability, each authority adopted equivalent national measures; in Spain, the CNMV Resolution of June 27, 2019, whose CFD chapter came into force on August 1, 2019 and still applies “as long as the CNMV does not consider that there have been changes in market conditions that allow its repeal” (unofficial translation).

flowchart TB
  A[MiFIR art. 40 and 42] --> B[ESMA 2018/796]
  B --> C[CNMV June 2019]
  C --> D[CNMV July 2023]

The 2019 resolution applies to all marketing, distribution and sale of CFDs to retail clients resident in Spain, whatever the country of the firm and whether or not it has a branch there; if another authority imposes something different, the stricter rule applies. It treats as CFDs, among others, rolling spot forex and spread bets, and it leaves out warrants and turbos. It does not apply to professional clients: that is why lesson 7.4.1 insists on what is lost by asking for that treatment.

The five conditions of Article 4

Article 4 of the resolution allows a CFD to be sold to a retail client only if all of these conditions are met:

Annex I sets the margin by type of underlying. The detail is in the lists: there are six “major currencies” (US dollar, euro, yen, pound, Canadian dollar and Swiss franc), and the 5% indices are ten named ones, with their 2019 names (“DAX30”). The IBEX 35 is not among them. A calculator that applies it as written:

import numpy as np
import pandas as pd

PRINCIPALES = {"USD", "EUR", "JPY", "GBP", "CAD", "CHF"}
INDICES_5 = {"FTSE 100", "CAC 40", "DAX30", "DJIA", "S&P 500",
             "NASDAQ", "NASDAQ 100", "Nikkei 225", "ASX 200",
             "EURO STOXX 50"}

def garantia_inicial(tipo, nombre=""):
    """Fraction of the notional (Annex I of the 2019 Resolution)."""
    if tipo == "divisa":
        a, b = nombre[:3], nombre[3:]
        return 0.0333 if {a, b} <= PRINCIPALES else 0.05
    if tipo == "indice":
        return 0.05 if nombre in INDICES_5 else 0.10
    if tipo == "materia_prima":
        return 0.05 if nombre == "oro" else 0.10
    if tipo == "cripto":
        return 0.50
    return 0.20                 # stocks and any other underlying

cartera = pd.DataFrame([
    ("divisa", "EURUSD", 100_000), ("divisa", "EURTRY", 20_000),
    ("indice", "DAX30", 40_000), ("indice", "IBEX 35", 40_000),
    ("materia_prima", "oro", 20_000),
    ("materia_prima", "brent", 20_000),
    ("accion", "una acción", 10_000), ("cripto", "BTC", 4_000),
], columns=["tipo", "subyacente", "nominal_eur"])
pares = zip(cartera["tipo"], cartera["subyacente"])
cartera["garantia_pct"] = [100 * garantia_inicial(t, s)
                           for t, s in pares]
cartera["garantia_eur"] = (cartera["nominal_eur"]
                           * cartera["garantia_pct"] / 100)
print(cartera.to_string(index=False))
total = cartera["garantia_eur"].sum()
nominal = cartera["nominal_eur"].sum()
print(f"total margin {total:,.0f} EUR for {nominal:,.0f} EUR of "
      f"notional (average leverage {nominal / total:.1f}:1)")

         tipo subyacente  nominal_eur  garantia_pct  garantia_eur
       divisa     EURUSD       100000          3.33        3330.0
       divisa     EURTRY        20000          5.00        1000.0
       indice      DAX30        40000          5.00        2000.0
       indice    IBEX 35        40000         10.00        4000.0
materia_prima        oro        20000          5.00        1000.0
materia_prima      brent        20000         10.00        2000.0
       accion una acción        10000         20.00        2000.0
       cripto        BTC         4000         50.00        2000.0
total margin 17,330 EUR for 254,000 EUR of notional (average leverage 14.7:1)

The data values of the table stay in Spanish, as in the rest of the course: divisa (currency), indice (index), materia_prima (commodity), accion (stock), oro (gold). The same notional in the DAX30 and in the IBEX 35 costs twice the margin in the second, and the EURTRY pair, with a currency that is not on the list, goes to 5%. The average leverage of the portfolio, which mixes bands, is none of those in the annex: that is what happens in a real account.

What maximum leverage means with data

ESMA did not pick the figures by eye. In its decision it explains that it calibrated the margin of each underlying according to its volatility, simulating with about ten years of daily prices the probability that a client loses 50% of their initial investment over an appropriate holding period, with the aim of making that probability consistent across underlyings. The same exercise can be done with the course data. The series imitate real markets; the band assigned to each is that of the market it imitates, and this is an assumption of the lesson:

The experiment: the whole account balance as margin (the maximum leverage allowed), one position opened every session, long or short, held for 20 sessions. It is closed out if at any daily close the loss reaches half of the margin, that is, an adverse move of half of the annex percentage:

from numpy.lib.stride_tricks import sliding_window_view

fx = pd.read_parquet("fx3_diario.parquet")
ind = pd.read_csv("futuros3_indices_contado.csv", parse_dates=["fecha"])
fut = pd.read_parquet("futuros2_continuos.parquet")
acc = pd.read_parquet("universo2_ajustados.parquet")
cri = pd.read_parquet("cripto3_diario.parquet")

def serie(df, col_id, valor, col_precio):
    s = df[df[col_id] == valor].set_index("fecha")[col_precio]
    return s.sort_index().dropna()

series = {
    "QEVUSD": (serie(fx, "par", "QEVUSD", "spot"), 0.0333),
    "USDQMP": (serie(fx, "par", "USDQMP", "spot"), 0.05),
    "RVE": (serie(ind, "indice", "RVE", "contado"), 0.05),
    "MEA": (serie(fut, "mercado", "MEA", "cierre_ajustado"), 0.05),
    "RVH": (serie(ind, "indice", "RVH", "contado"), 0.10),
    "ENA": (serie(fut, "mercado", "ENA", "cierre_ajustado"), 0.10),
    "XBQ": (serie(cri[cri["exchange"] == "boreal"], "activo", "XBQ",
                  "cierre_spot"), 0.50),
}

def prob_cierre(precios, garantia, h=20):
    """Fraction of h-session windows with a close-out (long/short mean)."""
    v = sliding_window_view(np.asarray(precios, float), h + 1)
    r = v[:, 1:] / v[:, :1] - 1
    largo = (r <= -0.5 * garantia).any(axis=1).mean()
    corto = (r >= 0.5 * garantia).any(axis=1).mean()
    return (largo + corto) / 2

res = {k: prob_cierre(s, g) for k, (s, g) in series.items()}
por_accion = [prob_cierre(g["cierre_ajustado"], 0.20)
              for _, g in acc.sort_values("fecha").groupby("ticker")
              if len(g) > 21]
res["stocks"] = float(np.mean(por_accion))
prob = pd.Series(res).sort_values(ascending=False)
(100 * prob).round(1)

RVE       52.7
ENA       51.6
MEA       49.8
QEVUSD    39.8
USDQMP    36.7
RVH       30.0
stocks    19.3
XBQ        7.4
dtype: float64

The figure at the top sorts the result. At maximum leverage with the whole balance committed, a CFD is closed out by the 50% rule in a large share of the four-week windows, and the share is not the same across underlyings. The single 5% band for major indices, gold and secondary currencies treats series with different volatilities alike, and the same happens between the IBEX 35 and crude oil at 10%. The cryptocurrency, with the 50% margin, is the one closed out least often: the band is conservative for its volatility. The figures depend on the synthetic data; what can be generalized is the method and the order of magnitude.

The way to reduce that probability is decided not by the rule but by the client: not committing the whole balance. The same account with less effective leverage on the major pair:

qev = series["QEVUSD"][0]
for apal in (30, 20, 10, 5, 2):
    # with effective leverage L, the close-out comes with a fall
    # of 50% of the balance: a move of 0.5 / L
    p = prob_cierre(qev, 1 / apal)
    print(f"{apal:>2}:1  move that closes out {50 / apal:5.2f} %  "
          f"prob. in 20 sessions {100 * p:5.1f} %")

30:1  move that closes out  1.67 %  prob. in 20 sessions  39.8 %
20:1  move that closes out  2.50 %  prob. in 20 sessions  23.0 %
10:1  move that closes out  5.00 %  prob. in 20 sessions   4.3 %
 5:1  move that closes out 10.00 %  prob. in 20 sessions   0.4 %
 2:1  move that closes out 25.00 %  prob. in 20 sessions   0.0 %

The simplification is honest in one sense: with the minimum margin and the whole balance committed, margin and balance coincide, and the close-out comes with the loss of half of the balance. With less leverage, the provider would not close out under the 50% rule (the balance is more than enough), but losing 50% of the balance is still the threshold that matters to the client.

The 50% close-out: per account, not per position

Article 3.e defines the protection as the liquidation of one or more CFDs “when the sum of funds in the CFD trading account and the unrealized net profits of all open CFDs connected to that account falls to less than half of the total initial margin provided for all those open CFDs” (unofficial translation of the CNMV resolution). It is a rule per account. ESMA explains in its decision that the public consultation proposed a rule per position and that it changed it: a rule per account allows the gains of one position to offset the losses of another, and a provider that closes out per position also complies, because it is stricter.

The difference matters as soon as there are two positions that hedge each other. A long-short pair: a long CFD on RVE (major index, 5%) and a short one on RVH (10%), with the same notional, and the account holding exactly the margin of both. The two series have a daily correlation of 0.73:

p = ind.pivot(index="fecha", columns="indice", values="contado")
par = p[["RVE", "RVH"]].to_numpy()

def ventana(i, regla, h=20):
    nominal, signo = 1.0, np.array([1, -1])     # long RVE, short RVH
    garantia = nominal * np.array([0.05, 0.10])
    fondos = garantia.sum()
    abierta = np.array([True, True])
    cerrada_pnl = np.zeros(2)
    cerro = False
    for t in range(1, h + 1):
        vivo = signo * nominal * (par[i + t] / par[i] - 1)
        pnl = np.where(abierta, vivo, cerrada_pnl)
        if regla == "position":
            corta = abierta & (pnl < -0.5 * garantia)
            cerrada_pnl[corta], abierta[corta] = pnl[corta], False
            cerro |= corta.any()
        if abierta.any() and (fondos + pnl.sum()
                              < 0.5 * garantia[abierta].sum()):
            cerrada_pnl[abierta], abierta[:] = pnl[abierta], False
            cerro = True
    return cerro, pnl.sum() / fondos

for regla in ("account", "position"):
    r = [ventana(i, regla) for i in range(len(par) - 20)]
    cerradas = np.mean([c for c, _ in r])
    final = np.array([x for _, x in r])
    print(f"{regla:9} close-outs {100 * cerradas:5.1f} %   mean "
          f"result {100 * final.mean():5.1f} %   5th percentile "
          f"{100 * np.percentile(final, 5):6.1f} %")

account   close-outs   6.6 %   mean result   0.9 %   5th percentile  -51.2 %
position  close-outs  71.8 %   mean result   2.0 %   5th percentile  -58.7 %

With the close-out per position, one leg is liquidated in most windows: the 5% leg withstands a 2.5% adverse move before being closed, and in a hedged pair each leg moves much more than the whole. The worst part is not the frequency, it is what is left: one leg open without its hedge, an exposure the client had not chosen. That is the argument ESMA takes from the public consultation. The rule per account looks at what the client can lose, the net, and barely acts. In an account with a single position, as ESMA says, the two rules are the same.

The CNMV adds a good practice in the preamble of the 2019 resolution: warn the client before the 50%, so they can post more margin or close on their own. And the 2023 resolution specifies that the close-out “must not take place before that threshold is reached and the client has been informed” (unofficial translation).

stateDiagram-v2
  [*] --> Normal
  Normal --> Warning: funds below threshold
  Warning --> Normal: adds funds or closes
  Warning --> CloseOut: below 50%
  CloseOut --> Normal: balance left
  CloseOut --> Protected: negative balance
  Protected --> [*]

Negative balance protection

The 50% close-out leaves the other half of the margin as a cushion to execute the close-out. That cushion is useless if the price gaps between two observations: a weekend, a central bank decision, a gap at the open. The currency data contain one like that. On Friday, October 2, 2015 QEVUSD closed at 1.58226 and on Monday the 5th, at 1.45688:

q = series["QEVUSD"][0]
salto = q.pct_change().abs().idxmax()
antes, despues = q.shift(1)[salto], q[salto]
mov = despues / antes - 1
fondos = 10_000                       # EUR in the CFD account
nominal = fondos / 0.0333             # the whole balance as margin
saldo = fondos + nominal * mov        # long CFD
print(f"{salto.date()}: {antes:.5f} -> {despues:.5f} "
      f"({100 * mov:.2f} %)")
print(f"notional {nominal:,.0f}; balance after the gap {saldo:,.0f} EUR")
print(f"client's loss with protection: {fondos:,.0f} EUR; "
      f"borne by the provider: {-min(saldo, 0):,.0f} EUR")

2015-10-05: 1.58226 -> 1.45688 (-7.92 %)
notional 300,300; balance after the gap -13,796 EUR
client's loss with protection: 10,000 EUR; borne by the provider: 13,796 EUR

For simplicity the account is in euros and the notional is expressed in euros. A 7.9% fall at 30:1 leverage is a loss of 238% of the margin: by the time the provider can close out, there is no balance left. Negative balance protection (Article 3.f) limits the client's liability, for all the CFDs in that account, to the funds deposited in it. How often this happens in the experiment of the previous section can be counted: the first close-out signal arrives with the loss already above 100% of the margin.

def saldo_negativo(precios, garantia, h=20):
    v = sliding_window_view(np.asarray(precios, float), h + 1)
    r = v[:, 1:] / v[:, :1] - 1
    out = 0
    for x in (r, -r):                          # long and short
        toca = x <= -0.5 * garantia
        hay = toca.any(axis=1)
        primero = toca.argmax(axis=1)
        perdida = -x[np.arange(len(x)), primero] / garantia
        out += int((hay & (perdida > 1)).sum())
    return out

pd.Series({k: saldo_negativo(s, g) for k, (s, g) in series.items()})

QEVUSD     88
USDQMP      4
RVE       167
MEA       116
RVH        16
ENA       232
XBQ         0
dtype: int64

These are windows (long and short) whose first close-out signal arrives with the loss already above the margin: 232 in crude oil, 167 in the major index, 88 in the major pair and none in the cryptocurrency, whose 50% margin leaves a cushion of a 25% move. Gaps are concentrated on a few dates and affect every window that spans them, which is why the counts are large in some series and small in others. With intraday data and continuous monitoring there would be fewer cases, not none: the gap between two prices exists at any frequency.

The risk warning and its percentage

Every communication from a CFD provider to retail clients carries a warning in a fixed format (Annex II). On a website or a durable medium (unofficial translation of the Spanish text):

CFDs are complex instruments and come with a high risk of losing money rapidly due to leverage. [provider's percentage] % of retail investor accounts lose money when trading CFDs with this provider. You should consider whether you understand how CFDs work and whether you can afford to take the high risk of losing your money.

The percentage is not free. Section A.6 defines it: every three months, over the previous 12 months, the proportion of retail CFD accounts whose net realized and unrealized result on CFDs was negative, with all costs (charges, fees, commissions) and without the accounts that had no open CFD, without the result of other products and without deposits or withdrawals. A provider with no open CFDs in the period uses the standard warning: “Between 74% and 89% of retail investors lose money when trading CFDs” (unofficial translation), the range from the supervisors' studies that the resolution itself cites.

A validator on 5,000 synthetic accounts, compared with two tempting and wrong calculations:

rng = np.random.default_rng(2019)
n = 5_000
cuentas = pd.DataFrame({
    "con_cfd_abierto": rng.random(n) < 0.85,
    "realizado": rng.normal(-150, 900, n).round(2),
    "no_realizado": rng.normal(0, 200, n).round(2),
    "costes": rng.gamma(2.0, 180, n).round(2),
    "otros_productos": rng.normal(300, 1500, n).round(2),
    "ingresos": rng.gamma(1.5, 2000, n).round(2),
    "retiradas": rng.gamma(1.2, 900, n).round(2),
})
cuentas.loc[~cuentas["con_cfd_abierto"],
            ["realizado", "no_realizado", "costes"]] = 0.0

def pct_pierden(c):
    """Annex II, section A.6 of the 2019 Resolution."""
    c = c[c["con_cfd_abierto"]]
    neto = c["realizado"] + c["no_realizado"] - c["costes"]
    return 100 * (neto < 0).mean()

c = cuentas
sin_costes = c[c["con_cfd_abierto"]]
mal_1 = 100 * ((sin_costes["realizado"] + sin_costes["no_realizado"])
               < 0).mean()
saldo = (c["realizado"] + c["no_realizado"] - c["costes"]
         + c["otros_productos"] + c["ingresos"] - c["retiradas"])
mal_2 = 100 * (saldo < 0).mean()
print(f"Annex II A.6:                       {pct_pierden(c):5.1f} %")
print(f"without costs:                      {mal_1:5.1f} %")
print(f"change in balance, all accounts:    {mal_2:5.1f} %")

Annex II A.6:                        70.3 %
without costs:                       56.0 %
change in balance, all accounts:     29.7 %

Both errors lower the figure. Forgetting costs flips the sign of accounts that make a small profit before commissions and financing. Using the change in balance of all accounts mixes in deposits, other products and accounts without CFDs, and turns into “winners” accounts that only received a transfer. The annex's method exists so that the figure is comparable across providers.

What the CNMV added in 2023 and what none of this covers

The CNMV Resolution of July 11, 2023, applicable since August 3, 2023, does two things. For CFDs it bans marketing communications aimed at retail clients in Spain, including those that redirect to a provider's website or form and public offers of free or token-price courses, seminars or demo accounts; sponsorship and brand advertising with that effect; and practices such as paying clients for bringing in clients, rewarding the sales network for client acquisitions or client losses, call centers that promote these products and deposits by credit card. And for other leveraged instruments (those whose maximum risk is not known at subscription or can exceed the initial outlay, such as futures or written options) it requires an initial margin and the 50% close-out for retail clients. The margin is the lower of two amounts: the one in the 2019 Annex I according to the underlying, or the one required by the trading venue; if the underlying is a crypto-asset that is not a MiFID financial instrument, the annex amount with no reduction.

esp = pd.read_csv("futuros2_especificaciones.csv")
ultimo = fut[fut["fecha"] == fut["fecha"].max()].set_index("mercado")
esp = esp.set_index("mercado").loc[["IDA", "MEA", "ENA", "DIB", "BOC"]]
esp["nominal"] = ultimo["cierre"] * esp["multiplicador"]
esp["centro_pct"] = 100 * esp["margen_inicial"] / esp["nominal"]
# assumptions: IDA as a major index; DIB, a non-major currency;
# a bond future is an "other underlying"
esp["anexo_pct"] = [5.0, 5.0, 10.0, 5.0, 20.0]
esp["exigible_pct"] = esp[["centro_pct", "anexo_pct"]].min(axis=1)
esp[["nombre", "nominal", "centro_pct", "anexo_pct",
     "exigible_pct"]].round(2)

                       nombre   nominal  centro_pct  anexo_pct  exigible_pct
mercado
IDA      Índice Aldebarán 500  389500.0        7.01        5.0          5.00
MEA                       Oro   41300.0        5.81        5.0          5.00
ENA              Crudo ligero   42670.0       12.89       10.0         10.00
DIB         Divisa Kestra/USD   60390.0        3.31        5.0          3.31
BOC            Bono a 10 años   95937.5        3.86       20.0          3.86

In three of the five level 2 futures (index, gold and crude oil) the annex figure is lower than the market margin and is the one required; in the currency and the bond the market margin applies, lower than the annex's 5% and 20%. The columns are centro_pct (trading venue margin), anexo_pct (Annex I) and exigible_pct (the one required). The resolution sets the minimum the provider must ask of the retail client. Its clearest practical effect is the 50% close-out, which a market's margin does not impose on the end client.

What none of these measures covers: professional clients; retail clients who trade with unauthorized firms (lesson 4.4.1); and market risk itself. They limit how much can be lost through leverage, not whether the strategy makes money.

Exercises

1. Repeat the experiment at the top with windows of 5 and 60 sessions. How does the ranking of the underlyings change? What does it tell you about the “appropriate holding period” ESMA used?

tabla = pd.DataFrame({
    h: {k: 100 * prob_cierre(s, g, h=h)
        for k, (s, g) in series.items()}
    for h in (5, 20, 60)
}).round(1)
tabla.columns = [f"{h} sessions" for h in tabla.columns]
tabla.sort_values("20 sessions", ascending=False)

        5 sessions  20 sessions  60 sessions
RVE           22.3         52.7         70.0
ENA           21.0         51.6         70.7
MEA           18.7         49.8         69.3
QEVUSD        10.8         39.8         62.5
USDQMP         8.1         36.7         59.8
RVH            6.1         30.0         54.9
XBQ            0.3          7.4         36.2

2. Modify ventana so that the account holds twice the funds of the required margin. How many close-outs are there now under each rule?

def ventana_holgada(i, regla, h=20, colchon=2.0):
    nominal, signo = 1.0, np.array([1, -1])
    garantia = nominal * np.array([0.05, 0.10])
    fondos = colchon * garantia.sum()
    abierta = np.array([True, True])
    cerrada_pnl, cerro = np.zeros(2), False
    for t in range(1, h + 1):
        vivo = signo * nominal * (par[i + t] / par[i] - 1)
        pnl = np.where(abierta, vivo, cerrada_pnl)
        if regla == "position":
            corta = abierta & (pnl < -0.5 * garantia)
            cerrada_pnl[corta], abierta[corta] = pnl[corta], False
            cerro |= corta.any()
        if abierta.any() and (fondos + pnl.sum()
                              < 0.5 * garantia[abierta].sum()):
            cerrada_pnl[abierta], abierta[:] = pnl[abierta], False
            cerro = True
    return cerro

for regla in ("account", "position"):
    c = np.mean([ventana_holgada(i, regla)
                 for i in range(len(par) - 20)])
    print(f"{regla:9} close-outs {100 * c:5.1f} %")

account   close-outs   0.0 %
position  close-outs  71.8 %

3. A provider advertises on a social network: “Trade the IBEX with 20:1 leverage and get a 100 EUR bonus”. Point out which parts of the 2019 and 2023 resolutions that ad breaches if it is aimed at retail clients in Spain.