Flow toxicity & maker-taker economics

M1 has been building one thing without stating it. Here it is, and every term is a lesson you have already done:

   MAKER’S EDGE PER UNIT TRADED

     +  spread captured        earn the half-spread            M0.4
     +  rebate                 exchange pays you to post       this lesson
     −  adverse selection      you traded with the informed    M1.2
     −  inventory cost         you’re holding risk (γqσ²τ)     M1.3
     −  fees & tech            colocation, feeds, clearing     M0.6 / M1.6
   ────────────────────────────────────────────────────────────────────
     =  what a market-making business actually optimises

Read that as a job description and the industry stops being mysterious. Nobody at these firms is trying to be fast, or to be tight, or to have volume. They are trying to make that expression positive, at scale, and every apparent quirk of their behaviour is one of those terms being managed.

Toxicity: adverse selection you can measure

Toxic flow is order flow that systematically loses money for the maker who fills it — the adverse-selection term, made operational. It’s the practitioner’s word for μ in Glosten–Milgrom.

The maker cannot see μ. But it can see something suggestive: order-flow imbalance. Informed flow is directional by construction; noise flow is roughly balanced. So sustained one-sidedness is evidence of information.

VPIN (volume-synchronised probability of informed trading) formalises that. Rather than clock time, bucket the tape into equal-volume buckets — so busy periods generate more observations, which is the point:

   for each bucket of V shares, classify volume as buy- or sell-initiated:

                  Σ over buckets  | V_buy − V_sell |
        VPIN  =  ───────────────────────────────────
                        n · V

   balanced flow  (V_buy ≈ V_sell)  →  VPIN → 0   → benign
   one-sided flow (all one way)     →  VPIN → 1   → toxic

Used live, a rising VPIN says widen, or stop quoting. Easley, López de Prado and O’Hara showed it climbing sharply before the May 2010 Flash Crash, and argued it works as an early-warning signal for liquidity withdrawal.

It is contested, and you should know that. Andersen and Bondarenko argued the result is largely an artefact — that VPIN is substantially a repackaging of volatility and volume, that its bulk volume classification is noisy, and that its predictive claims don’t survive careful testing. The honest position: the concept of toxicity is indispensable and every real maker estimates something like it; VPIN specifically is one contested estimator, not a settled measurement. Say that in an interview and you’ll sound like someone who read past the abstract.

Maker-taker: exchanges paying for liquidity

Venues compete for order flow, and liquidity is self-reinforcing — a book with depth attracts more depth. So exchanges pay for it.

   MAKER-TAKER  (the common model)
     post a resting order that gets filled  →  you RECEIVE a rebate
     take liquidity with a marketable order →  you PAY a fee
     typical US equities: −0.20 ¢ / +0.30 ¢ per 100 shares
     the exchange keeps the difference

   TAKER-MAKER / “inverted”  (some venues)
     exactly reversed — pays takers, charges makers.
     Exists to attract takers to a thin book, and because some
     flow is worth paying for (see segmentation, below).

Two consequences that matter more than the arithmetic:

The economically real spread is net of rebates. A quoted spread of one cent, with a rebate of 0.2 cents per side, is not a one-cent spread to the participants. It can be rational to post at a price that looks like it loses money on the spread alone, because the rebate makes it profitable — which is precisely how markets stay quoted at the minimum tick.

Routing incentives and client interests come apart. The rebate accrues to whoever posts the order — the broker — while best execution is owed to the client. A broker choosing between venues offering identical prices but different rebates has an interest that is not the client’s. This is a live regulatory question, it is why Reg NMS’s protection of displayed prices doesn’t settle the matter, and it is the sharpest structural criticism of maker-taker.

Segmentation: the punchline of M0 and M1 together

Now put PFOF (M0.3), adverse selection (M1.2) and toxicity together, because the combination explains the shape of the modern US equity market.

   ALL ORDER FLOW
        │
        ├──▶ RETAIL ──▶ wholesaler / internaliser
        │              low μ, predictable, benign
        │              → filled off-exchange, INSIDE the quoted spread
        │              → wholesaler pays the broker for it (PFOF)
        │
        └──▶ EVERYTHING ELSE ──▶ the lit book
                       institutions, funds, other makers
                       → higher average μ  →  MORE TOXIC
                       → lit spreads must be WIDER to survive it

Wholesalers aren’t being generous when they price-improve retail orders. They are paying for the right kind of counterparty — and every benign order they remove raises the average toxicity of what’s left on the lit book, which must widen accordingly.

That gives you the genuinely hard policy question, which has no clean answer: retail investors demonstrably get better prices than the displayed quote, while the displayed quote itself is worse than it would be if their flow were in it. Whether that is good depends on who you weigh, and anyone confident in either direction is skipping a step.

What this module was for

You can now read a real market as a system of forces rather than a set of terms:

• A spread is order-processing plus inventory plus adverse selection (M0.4), and you can derive the last two (M1.2, M1.3).
• A price move is permanent or temporary, and which one it is tells you whether it was information or inventory (M1.5).
• Depth is not resting size — it is how hard it is to distinguish information from noise (M1.4).
• Speed buys queue position and stale quotes, and nothing else (M1.6).
• A market-making business is the equation at the top of this page (M1.7).

M2 changes the subject to pricing and strategy, and M5 returns to these same firms to ask how they’re actually organised and paid. But the layer in this module is the one most of them live in.

Source: Easley, López de Prado & O’Hara (2012), “Flow Toxicity and Liquidity in a High-Frequency World” for VPIN — then Andersen & Bondarenko (2014), “VPIN and the Flash Crash”, for the rebuttal; reading them in that order is the point. On fee structures, the SEC’s Transaction Fee Pilot materials are the most concrete public source. O’Hara, “High-frequency market microstructure” (2015) is a good survey of the whole modern layer.