Market participants and trading venues.

An order passes through many hands

When a retail investor taps “buy” in their broker's app, the order does not go straight “to the exchange”. It travels through a chain of intermediaries, and each one adds a function, a risk and a cost. Knowing the chain explains many of the odd things that show up in the data: trades at half-cent prices, volume that is not in the book, prices that arrive out of order.

flowchart TB
  I[Investor] --> B[Broker]
  B --> S[SOR router]
  B --> SI[Internalizer]
  S --> R[Exchange]
  S --> M[MTF]
  S --> O[Dark pool]
  R --> C[CCP clearing house]
  M --> C
  O --> C
  C --> D[CSD depository]

• Investor. Decides what to buy or sell. It can be a person, a fund or a program.
• Broker. The authorized firm that receives the order, checks that there is enough cash or collateral and sends it for execution. In the EU it is an investment firm or a bank; it is obliged to seek best execution for its client (price, cost, speed, likelihood of execution).
• Router (SOR, smart order router). The broker's program that decides which venue to send the order to, or how to split it among several, depending on where the best price and liquidity are at that moment.
• Execution venues. The places where the order is matched with someone else's: exchanges, MTFs, dark pools (all multilateral) or an internalizer (bilateral). The next section tells them apart.
• Clearing house (CCP, central counterparty). After the trade it steps in between buyer and seller: it becomes the buyer to every seller and the seller to every buyer. If one party defaults, the clearing house performs. To cover that risk it requires collateral (margins) from its members.
• Central securities depository (CSD). Keeps the record of who holds each security and delivers against payment: settlement, one or two business days later (lesson 1.1.5).

Who is on the other side

Every purchase has a sale on the other side. Knowing who is usually on the other side helps you understand why a market behaves the way it does. The figure at the top places the participants on two logarithmic axes: how long they hold a position and how large their orders are. These are orders of magnitude, not measurements, but the distances are real: between a market maker and a pension fund there are twelve orders of magnitude of horizon.

Two consequences for a trading system. First: almost all immediate liquidity is provided by market makers who do not want to hold positions; if their risk rises (volatility, news), they pull their orders and the spread widens (lesson 1.1.3). Second: a low-frequency system competes on its horizon with asset managers and CTAs, not with HFT. Speed is neither its edge nor its problem; costs and look-ahead bias are.

Where trading happens: exchanges, MTFs, dark pools and internalizers

European regulation (MiFID II and MiFIR) recognizes three types of multilateral trading venue, where many buyers and sellers are matched with each other under non-discretionary rules, and one bilateral figure, the internalizer. Outside all of them lies OTC.

In Europe a “dark pool” is not a separate legal type: it is a regulated market or an MTF that operates with a pre-trade transparency waiver, for example matching at the reference price (the midpoint of another venue) or matching large-in-scale orders. Since those trades do not contribute to price formation, regulation limits how much volume can be traded that way.

In the US the names change (registered exchanges, ATSs for dark pools, wholesalers that internalize retail orders), but the structure is the same: lit books fragmented across many exchanges and a large share of the volume outside them.

A NOVA session seen by venue

The data for this lesson is a full session of NOVA, a fictitious stock listed in the US, on June 2, 2025: 50,946 trades with their price, size, aggressor side and execution venue. It comes from a book simulated order by order, so trades, quotes and book are consistent with each other. The data is synthetic, but it follows real conventions: a one-cent tick, a session from 9:30 to 16:00 New York time (13:30 to 20:00 UTC) and opening and closing auctions.

import pandas as pd

op = pd.read_parquet("nova_operaciones.parquet")
print(len(op), "trades")
op.head(4)

50946 trades
                                ts  precio  tamano agresor     centro         condicion
0        2025-06-02 13:30:00+00:00   53.35   33900     NaN  PRINCIPAL  subasta_apertura
1 2025-06-02 13:30:00.301051+00:00   53.34     100   venta  PRINCIPAL           regular
2 2025-06-02 13:30:00.301052+00:00   53.34     100   venta  PRINCIPAL           regular
3 2025-06-02 13:30:00.407366+00:00   53.34     159   venta  PRINCIPAL           regular

Each row is a trade. agresor says which side took the initiative: the order that arrived and was executed against one already waiting in the book. In the opening auction it is empty, because an auction has no aggressor (lesson 1.1.2). Grouping by venue:

por_centro = op.groupby("centro").agg(
    operaciones=("tamano", "size"),
    volumen=("tamano", "sum"),
    tamano_medio=("tamano", "mean"),
)
total = por_centro["volumen"].sum()
por_centro["cuota_pct"] = 100 * por_centro["volumen"] / total
por_centro.sort_values("volumen", ascending=False).round(1)

                operaciones  volumen  tamano_medio  cuota_pct
centro
PRINCIPAL             25776  3564476         138.3       40.1
OSCURO                 4386  2358700         537.8       26.5
INTERNALIZADOR         3914  1243861         317.8       14.0
MTF_A                  9640   986282         102.3       11.1
MTF_B                  7230   735395         101.7        8.3

The primary venue, the one with the reference book and the auctions, gathers 40% of the volume. The two MTFs add almost another 20%. And a quarter of the volume is matched in a dark pool: nobody saw it in a book before it happened. NOVA's figures come from a simulation, but the pattern —no venue concentrates all the volume— is that of any liquid stock in the US or in Europe.

The average size tells the other half of the story:

In the venues with a book, a typical trade is about 100 shares: orders executed against what sits at the first level, many of them slices of a larger order split by an algorithm (lesson 1.1.4). In the dark pool, the average size is five times larger. The reason is economic: a large order shown in a lit book moves the price before it is executed, because the others step aside or front-run it. Matching it blind at the midpoint avoids that.

The condicion column separates the special trades:

op.pivot_table(index="centro", columns="condicion",
               values="tamano", aggfunc="sum", fill_value=0)

condicion       bloque  regular  subasta_apertura  subasta_cierre
centro
INTERNALIZADOR       0  1243861                 0               0
MTF_A                0   986282                 0               0
MTF_B                0   735395                 0               0
OSCURO          324900  2033800                 0               0
PRINCIPAL            0  2645076             33900          885500

Auctions only exist in the primary venue: the opening auction matches little volume and the closing one almost 900,000 shares at a single price, because index funds and many asset managers execute at the close, which is the price they are valued at (lesson 1.1.2). The blocks, all in the dark pool, are a few huge trades:

bloques = op[op["condicion"] == "bloque"]
bloques[["ts", "precio", "tamano", "centro"]]

                                    ts  precio  tamano  centro
5478  2025-06-02 13:50:31.017740+00:00  53.375   64100  OSCURO
24035 2025-06-02 16:22:35.166309+00:00  53.795   46100  OSCURO
33823 2025-06-02 17:54:43.496329+00:00  53.255   63000  OSCURO
37493 2025-06-02 18:28:54.725977+00:00  53.195   36800  OSCURO
41959 2025-06-02 19:05:40.906104+00:00  52.655   68200  OSCURO
47010 2025-06-02 19:40:27.200952+00:00  52.570   46700  OSCURO

Six trades that add up to 324,900 shares, 3.7% of the day's volume. A rule that filters “abnormal volume” without looking at the condition would take these blocks for signals; they are not: they are institutions moving positions through negotiated trades.

What the dark pool does not show

The dark pool has no visible book, but it does have a price rule: it matches at the midpoint of the best bid and the best ask of the lit book. Checking it requires joining each trade with the quote prevailing at that instant. The quotes (nova_cotizaciones.parquet) are the best bid and the best ask of the lit book, one row every time they change:

cot = pd.read_parquet("nova_cotizaciones.parquet")
print(len(cot), "quote changes")

# the prevailing quote is the last one BEFORE the trade
con_cot = pd.merge_asof(op, cot, on="ts", allow_exact_matches=False)
con_cot["medio"] = (con_cot["bid"] + con_cot["ask"]) / 2
con_cot.loc[1:3, ["ts", "precio", "agresor", "centro", "bid", "ask"]]

129247 quote changes
                                ts  precio agresor     centro    bid    ask
1 2025-06-02 13:30:00.301051+00:00   53.34   venta  PRINCIPAL  53.34  53.36
2 2025-06-02 13:30:00.301052+00:00   53.34   venta  PRINCIPAL  53.34  53.36
3 2025-06-02 13:30:00.407366+00:00   53.34   venta  PRINCIPAL  53.34  53.36

merge_asof is a nearest-key join: for each trade it looks for the last quote with an earlier timestamp. The detail allow_exact_matches=False matters. When an aggressive order sweeps a level of the book, the quote changes in the same microsecond as the trade. If the exact match is allowed, that trade is joined with the quote it caused itself: information from the next instant. It is the look-ahead bias of lesson 0.1.1, now at microsecond scale:

def ejecuta_al_mejor_precio(df):
    con_libro = df[df["centro"].isin(["PRINCIPAL", "MTF_A", "MTF_B"])
                   & (df["condicion"] == "regular")]
    lado, precio = con_libro["agresor"], con_libro["precio"]
    compra = (lado == "compra") & (precio >= con_libro["ask"])
    venta = (lado == "venta") & (precio <= con_libro["bid"])
    return (compra | venta).mean()

mal = pd.merge_asof(op, cot, on="ts")  # allow_exact_matches=True
print(f"With exact matches: {ejecuta_al_mejor_precio(mal):.2%}")
print(f"Without them: {ejecuta_al_mejor_precio(con_cot):.2%}")

With exact matches: 99.91%
Without them: 100.00%

With the correct alignment, every aggressive buy pays the ask or more and every sell receives the bid or less, which is how a book works. With the wrong one, a few dozen trades seem to execute at a better price than was available. In an execution cost study (lesson 1.1.5) that error distorts the effective spread.

With the quote properly aligned, the rule of the dark pool and that of the internalizer can be checked:

oscuro = con_cot[con_cot["centro"] == "OSCURO"]
al_medio = (oscuro["precio"] - oscuro["medio"]).abs() < 1e-9
print(f"Dark crosses at the midpoint: {al_medio.mean():.1%}")

inter = con_cot[con_cot["centro"] == "INTERNALIZADOR"]
mejora = (inter["ask"] - inter["precio"]).where(
    inter["agresor"] == "compra", inter["precio"] - inter["bid"])
print(f"Internalizer with price improvement: {(mejora > 0).mean():.1%}")
print(f"Average improvement: {100 * mejora.mean():.2f} cents per share")

Dark crosses at the midpoint: 100.0%
Internalizer with price improvement: 44.1%
Average improvement: 0.24 cents per share

All the dark crosses are at the midpoint, as their rule says. The internalizer executes against its own account inside the spread and, in a little under half of the trades, improves on the price of the lit book: that is what it offers in exchange for keeping the order flow. For the client, half a cent of improvement on a $53 stock is little; for the internalizer, multiplied by millions of orders, it is a business.

flowchart TB
  O[Client order] --> T{Is it large?}
  T -- yes --> OS[Dark pool or block]
  T -- no --> P{Price improvement?}
  P -- yes --> IN[Internalizer]
  P -- no --> SOR[Lit books]

The diagram simplifies the decision of the router and the broker. In practice it is made by algorithms with data from each venue, and the best execution obligation requires being able to justify it.

Clearing and settlement

The trade is only the beginning. Two steps follow that the investor barely sees and that explain some of the data:

flowchart LR
  N[Trading] --> C[CCP clearing] --> L[CSD settlement]

• Clearing. The CCP computes each member's net positions (many buys and sells of the same security cancel out) and requires collateral for the outstanding risk.
• Settlement. The CSD delivers the securities and the cash. In the US it happens on the next business day (T+1) since May 28, 2024; in the EU, two days later (T+2), with the move to T+1 scheduled for October 11, 2027.

For a trading system, settlement matters in two places: the cash from a sale is not available until it settles (in a cash account), and the date that determines who receives a dividend is the settlement date, which is why the ex date exists (lessons 1.1.5 and 1.2.1).

What this changes in your data

• Closing price ≠ last trade. The official close is that of the primary venue's auction; later trades on other venues or after hours do not change it.
• The day's volume is split. A provider that only gives the primary exchange's volume (some do) shows 40% of NOVA's real volume. Before comparing volumes, you need to know which tape each source uses (lesson 1.3.1).
• Not all trades are equal. Auctions, blocks and after-hours trades carry condition codes. An indicator such as VWAP (lesson 1.7.1) gives different results with and without them.
• The book is not all the liquidity. What you see at the first level is part of the real interest; the dark pool and the internalizers match without going through it.

Exercises

1. Compute each venue's volume share using only the trades with condition regular, without auctions or blocks. Which venue gains weight and why?

regular = op[op["condicion"] == "regular"]
por_centro_reg = regular.groupby("centro")["tamano"].sum()
cuota = 100 * por_centro_reg / regular["tamano"].sum()
cuota.sort_values(ascending=False).round(1)

centro
PRINCIPAL         34.6
OSCURO            26.6
INTERNALIZADOR    16.3
MTF_A             12.9
MTF_B              9.6
Name: tamano, dtype: float64

2. Compare the mean and the median trade size of each venue. What does the difference between the two tell you in the dark pool?

op.groupby("centro")["tamano"].agg(["mean", "median"]).round(0)

                 mean  median
centro
INTERNALIZADOR  318.0   236.0
MTF_A           102.0   100.0
MTF_B           102.0   100.0
OSCURO          538.0   300.0
PRINCIPAL       138.0   100.0

3. A system buys 2,000 NOVA shares with a market order at 15:00 New York time. Explain on which venues it could be executed, who would probably be on the other side and which data from this lesson you would use to estimate its cost.