Market Impact as Causal TCA

For six lessons you have been hunting causal effects you can never fully observe: does the signal cause the return, or merely co-move with it? You squinted at backtests, drew DAGs, and begged the market for natural experiments. This lesson hands you the one causal effect you do not have to beg for — because you produce it yourself, every single time you trade. When you buy, the price tends to rise; when you sell, it tends to fall. That move is not someone else’s alpha leaking out. It is your own footprint, your treatment, your effect. Market impact is the cleanest causal question in all of finance, and — in the cruellest twist of the course — measuring it honestly is still confounded, because you do not trade at random. You stomp hardest exactly when the floor was already shaking.

This is the capstone: causal inference turned on the one experiment you are guaranteed to run, transaction-cost analysis (TCA) done as a treatment-effect estimate instead of a naive regression, and finally — having taught you all the machinery — an honest accounting of what causal inference in markets cannot deliver.

Market impact is a causal effect

Analogy. Step into a still pond and the water rises around your ankles. The cost of getting wet is not the pond’s total level — it is how much higher the water sits because you are standing in it. To bill yourself honestly you would need the water level in a parallel pond you never entered. You never get that pond. You only ever observe the one you are standing in, ripples and all.

Definition. Implementation shortfall is the gap between the price at your decision and the average price you actually realised, including the part of the move your own order caused. Split it into two pieces:

• Temporary impact — the part that comes from demanding immediacy: you consume the order book’s near-side liquidity, the price ticks against you while you trade, and it relaxes back once you stop. This is a rent you pay for speed.
• Permanent impact — the part that stays after you finish, because the market reads your trade as information (“someone knows something”) and re-prices accordingly.

Formally, the causal cost of your order is the realised price path minus the counterfactual price path had you not traded. That counterfactual is unobservable — it is the fundamental problem of causal inference (one unit, only one of its two potential outcomes is ever seen) wearing an execution costume.

Worked example. You buy 100,000 shares; the decision price is $50.00. You realise an average fill of $50.05, and after you finish the price settles at $50.02 (the rest of the market having drifted nowhere on its own).

• Total realised cost per share: $50.05 - 50.00 = 0.05$ dollars, i.e. $0.05 / 50.00 = 10$ bps.
• Permanent impact: the price stuck at $50.02$, so $50.02 - 50.00 = 0.02$ dollars, which is $4$ bps that stayed.
• Temporary impact: the rest, $10 - 4 = 6$ bps, relaxed away after you stopped pressing.

Dollar cost: $100000 \times 0.05 = 5000$, so $5000 of implementation shortfall, of which roughly $2000 is permanent and $3000 is the temporary rent for trading fast.

When to use it

Reach for the treatment-effect framing whenever you are about to attribute a cost or a benefit to your own action — sizing a block, choosing an algo, judging a broker. The moment the question is “what did my trading do to the price?”, you are estimating a causal effect with an unobservable counterfactual, and you should reach for the identification toolkit, not a spreadsheet of raw averages.

Why naive TCA is confounded

Analogy. Imagine measuring whether umbrellas cause rain by noting that on every day people carried umbrellas, it rained. The umbrella didn’t summon the clouds — people chose to carry one because the forecast was already grim. Naive TCA does the same thing in reverse: you “carry the big order” (trade aggressively) precisely on the days the price was already going your way. Blame the order for the whole move and you have credited the umbrella with the storm.

Definition. Naive TCA regresses realised cost on observed order characteristics (size, participation rate, urgency) as if those characteristics were assigned. They are not — they are chosen by a trader reacting to the same information that drives the price. The trading decision is a confounder: it causally influences both the treatment (how big and fast you trade) and the outcome (the price move). The clean back-door story from earlier lessons applies exactly: size $\leftarrow$ signal $\rightarrow$ return opens a non-causal path, so the regression coefficient mixes (a) the genuine causal impact of trading with (b) the alpha/momentum you selected on.

Worked example. Decompose the realised cost on aggressive versus passive orders. Suppose the true causal impact of going aggressive is +6 bps. But aggressive orders were *fired specifically when momentum was already pushing +5 bps in your favour during the trade — favourable drift you would have captured regardless. Passive orders were used on quiet, driftless names.

The naive comparison says aggressive trading is cheaper (−1 bps) when its true causal impact is more expensive (+4 bps). The confounder didn’t merely inflate the estimate — it flipped its sign. A desk trusting this would conclude “trade more aggressively, it’s free,” and proceed to set fire to its execution budget.

When to use it

Treat every raw cost-versus-size or cost-versus-urgency number as confounded until proven otherwise — especially flattering ones (“our aggressive orders are cheap!”). Before acting on a TCA regression, ask the lesson-3 question: what decided the order size, and does that same thing move the price? If the answer is “our own alpha,” you are looking at selection, and you need an identification strategy, not a fitted line.

The square-root law of impact

Analogy. Pouring sand into a bucket of water: the first scoop raises the level the most, the tenth scoop barely registers because the water has spread out to accommodate it. Liquidity behaves the same way — the market absorbs additional size with diminishing marginal pain. Impact is concave, not a straight ramp.

Definition. The square-root law of market impact, the most robust empirical regularity in execution, models expected impact cost as

where $\sigma$ is the asset’s volatility, $Q$ is the order size, $V$ is average daily volume (ADV), and $Y$ is a dimensionless constant of order one. The $\sqrt{\,\cdot\,}$ is the punchline: cost grows with the square root of participation, so it is concave — steep for the first percent of volume, then flattening. Crucially, this is a relationship people work hard to estimate causally (from controlled or quasi-controlled variation in $Q$), not a naive regression of cost on whatever sizes traders happened to choose.

Worked example. Take $\sigma = 2\%$ daily volatility (200 bps), $Y = 0.5$, and you trade $Q/V = 10\%$ of ADV.

• $\sqrt{0.10} \approx 0.316$.
• Impact $\approx 0.5 \times 200 \text{ bps} \times 0.316 \approx 31.6$ bps.

Now double the order to $Q/V = 20\%$:

• $\sqrt{0.20} \approx 0.447$.
• Impact $\approx 0.5 \times 200 \times 0.447 \approx 44.7$ bps.

Doubling the order raised cost per share by a factor of only $0.447 / 0.316 \approx 1.41$ — that is $\sqrt{2}$, not $2$. The total dollar cost still rises (you are trading twice the shares at 1.41x the rate), but the per-share penalty grows sub-linearly. That concavity is exactly why working a large order in slices, and why strategy capacity, both scale better than the naive linear intuition fears.

When to use it

Use the square-root law as the default impact model for pre-trade cost estimates, capacity analysis, and the impact term inside an execution optimiser. Lean on its shape with confidence — concavity is one of the few near-universal facts in microstructure. Distrust its coefficient until you can show $Y$ was calibrated from size variation that wasn’t itself driven by alpha; otherwise you have baked the confounder into the model you optimise against.

Identifying impact cleanly

Analogy. Recall the whole arc of this course. You wanted a coin-flip to decide who got the treatment; markets almost never flip the coin for you, so you learned to manufacture as-good-as-random variation — instruments, discontinuities, natural experiments. Execution is the one corner of finance where you can sometimes flip the coin yourself: route orders at random and watch what the price does. When you can’t, you fall back on every quasi-experimental trick from lessons 4 through 6.

Definition. A ladder of identification strategies for impact, cleanest first:

• Randomised algo wheel (a real RCT). Take a stream of statistically similar parent orders and randomly assign each to broker/algo A or B (or to a higher/lower participation rate). Random assignment makes the trading decision independent of alpha and momentum, so the average cost difference is the unconfounded causal effect. This is the luxury the rest of the course kept telling you markets withhold — and execution is where you occasionally get it.
• Instrumented order size. When you cannot randomise, find an instrument (lesson 5): something that shifts your order size but does not touch the price except through that size. A risk-limit cap or a fund-flow-driven rebalance can move how much you must trade for reasons unrelated to your alpha.
• Natural experiments / exogenous flow. Scheduled index reconstitutions, passive-fund creation/redemption, and dividend reinvestment force trades on a calendar, with sizes set by index weights rather than by anyone’s view — quasi-random size variation (lesson 4).
• Difference-in-differences across venues. When a rule, fee, or tick-size change hits one venue and not another, compare the cost change on the treated venue to the change on an untreated one (lesson 6), differencing out common market-wide moves.

Worked example — the algo wheel as an A/B test. Route 1000 statistically matched orders, 500 to broker A and 500 to broker B at random. A’s average shortfall is 7.5 bps; B’s is 9.0 bps. Because assignment was random, the 1.5 bps difference is causal — no need to worry that A “got the easy orders,” since the coin flip guaranteed both got the same mix. Compare that to the naive alternative: brokers self-select which orders they pitch for, so a non-randomised comparison would tell you which broker is the better salesman, not the better executor.

When to use it

Default to the highest rung of the ladder your governance allows. If compliance and your brokers permit an algo wheel, run one continuously — it is the rare free RCT and it compounds into the best impact data you will ever own. When you cannot randomise, reach down the ladder to an instrument, a forced-flow natural experiment, or a venue DiD, and be explicit about the assumption each one rests on. The further down the ladder, the louder you should state the assumption you are buying.

The link to optimal execution — and the honest limits

Analogy. Optimal execution is a thermostat: trade too fast and you cook the order in impact; trade too slow and you freeze, exposed to the price wandering off (timing risk). The thermostat’s setting is your impact estimate. Feed it a miscalibrated thermometer — one biased by selection — and it will hold the room at confidently the wrong temperature, certain it is comfortable while everyone sweats.

Definition. Optimal execution (the Almgren-Chriss tradition) chooses a trading schedule to minimise expected impact cost plus a risk penalty on timing risk: trade quickly and you pay more impact but carry less exposure to adverse price moves; trade slowly and you cut impact but sit longer in the market’s path. The impact coefficient — your calibrated $Y$ from the square-root law — is the pivotal input. So every confounding sin from the earlier sections flows straight downstream: a biased impact estimate yields a biased trade-off, and you systematically over- or under-trade.

Worked example. Two desks calibrate the same square-root law on the same trades. Desk A naively fits $Y$ on alpha-selected orders and, because favourable drift was credited to small orders, understates impact, landing at $Y = 0.3$. Desk B uses algo-wheel-randomised data and recovers the honest $Y = 0.5$.

• Desk A’s optimiser sees cheap impact, so it front-loads aggressively — say targeting 20% of ADV per interval. Real impact at that pace (true $Y = 0.5$) is $0.5 \times 200 \times \sqrt{0.20} \approx 44.7$ bps, far more than the $\approx 26.8$ bps it budgeted.
• Desk B, with the honest coefficient, schedules a gentler 10% pace, pays $\approx 31.6$ bps, and avoids overpaying for speed it didn’t need.

Same market, same law, same shape — but a confounded coefficient led Desk A to trade too fast and overspend by roughly 13 bps on the aggressive slices. The bias didn’t stay in the spreadsheet; it walked into the market and cost real money.

Now the close. You have a complete toolkit — potential outcomes, DAGs, confounding, instruments, discontinuities, difference-in-differences, and TCA as a treatment effect. It would be dishonest to send you off thinking it proves anything. It does not. Here are the limits, stated plainly:

When to use it

Treat causal inference as a tool for sizing your conviction, not for certifying truth. Use it before every allocation and every execution decision to ask “is this a treatment effect or a back-door artefact, and how would I know?” — then act in proportion to how many independent, differently-flawed designs agree. Reserve your largest bets and your most aggressive schedules for effects that survive triangulation; haircut everything that rests on a single untestable assumption. The discipline is not certainty. It is knowing exactly which assumption you are betting the book on, and pricing it.

Recap

You arrived at the one causal effect you cannot dodge: trade, and you move the price. Implementation shortfall is that footprint — temporary impact (the rent for immediacy that relaxes back) plus permanent impact (the information the market reads into your order) — measured against a counterfactual price you never get to see. And in the course’s final twist, even this clean causal question is confounded: you size up and speed up exactly when alpha or momentum is already in your favour, so naive cost-on-size TCA mixes your real footprint with the drift you selected into — and can flip sign entirely. The square-root law gives you a robust concave shape but a coefficient that absorbs selection unless you calibrate it from exogenous size variation. The cleanest fix is the rare market RCT — an algo wheel — with instruments, forced-flow natural experiments, and venue DiD as the fallback ladder. Get the impact coefficient wrong and the optimal-execution schedule trades too fast or too slow with total confidence. And underneath all of it: no observational design proves anything, because the counterfactual is forever unseen. Causal inference in markets is the art of triangulating imperfect designs into a defensible degree of belief — and knowing precisely which untestable assumption you are betting on.