The Synthetic Control Method

A single exchange changes its tick-size rules and an analyst wants to know the effect on trading volume. There's no other exchange that's a close match on its own — one is too small, another trades a different mix of instruments, a third has different market hours. No individual comparison exchange is a fair control. The synthetic control method's insight: don't pick one imperfect comparison, build a custom blend of several imperfect comparisons, weighted so the blend's pre-change history tracks the treated exchange closely — and let that blend's post-change path stand in for what would have happened without the rule change.

An analogy: mixing a paint color you don't have

You need a specific shade of green paint but the store only sells blue, yellow, and a slightly-off green. Instead of using any one can as an imperfect substitute, you mix a custom blend — 40% blue, 35% yellow, 25% off-green — that matches your target shade as closely as possible when tested on a swatch you can already see (the "pre-treatment" period). Once you're confident the blend matches on the swatch you can check, you trust the same blend recipe to tell you what the color would look like on a wall you haven't painted yet (the "post-treatment" period) — the mixing weights, not any single can, are what does the work.

The method, one symbol at a time

Let $Y_{1t}$ be the treated unit's outcome at time $t$, and $Y_{jt}$ the outcome for each of $J$ untreated "donor" units. Choose non-negative weights $w_j$ summing to 1 that minimize the gap between the treated unit and the weighted blend during the pre-treatment period $t < T_0$:

$$\min_{w_1,\dots,w_J \ge 0,\; \sum w_j = 1} \sum_{t < T_0} \left( Y_{1t} - \sum_{j=1}^{J} w_j Y_{jt} \right)^2$$

In plain English: find the combination of donor units — using only nonnegative weights that add to one, so it's a genuine weighted average, never an extrapolation beyond the data — that best reproduces the treated unit's own history before the change. The estimated treatment effect at any later time $t \ge T_0$ is simply:

$$\hat{\tau}_t = Y_{1t} - \sum_{j=1}^{J} w_j Y_{jt}$$

the treated unit's actual outcome minus what the synthetic blend (extended forward using the same weights) would predict.

Worked example 1: constructing the blend

Suppose three donor exchanges have pre-change monthly volumes (in $bn) over three months: Donor A: 10, 11, 12. Donor B: 6, 5, 7. Donor C: 20, 19, 21. The treated exchange: 12.8, 12.7, 13.9. Solving the weighting problem might yield $w_A = 0.5$, $w_B = 0.3$, $w_C = 0.2$ (illustrative, found by minimizing squared pre-period gaps): synthetic path $= 0.5(10,11,12) + 0.3(6,5,7) + 0.2(20,19,21) = (5+1.8+4, 5.5+1.5+3.8, 6+2.1+4.2) = (10.8, 10.8, 12.2)$ — reasonably close to but not exactly matching (12.8, 12.7, 13.9); in practice you'd refine weights further, but this illustrates the mechanics: the blend tracks the shape of the treated unit's history using only real donor data.

Worked example 2: reading the post-treatment gap

Suppose after refinement the synthetic blend tracks the treated exchange near-exactly pre-change (gap under 0.1 every month), and the tick-size rule changes in month 4. Post-change: treated exchange volume is 11.5, while the synthetic blend (same weights, extended using donors' actual month-4 volumes) is 13.6. The estimated effect is $\hat\tau_4 = 11.5 - 13.6 = -2.1$, i.e., roughly a $2.1bn reduction in monthly volume attributable to the rule change. Because the pre-period fit was tight, this gap is much more credible than it would be if the synthetic blend had only loosely tracked the treated exchange beforehand — the tightness of the pre-period fit is itself the evidence for (or against) trusting the post-period gap.

What this means in practice

Synthetic control is the standard tool when you have exactly one (or very few) treated units and a panel of untreated candidates, none of which alone is a convincing match — a common situation for exchange-level or country-level policy changes where randomization was never possible and there's no natural single control group. It shares difference-in-differences' reliance on a parallel-trends-style assumption, but instead of assuming it holds for one chosen control, it constructs a control specifically engineered to satisfy it, and the tightness of the pre-period fit is a direct, visible diagnostic of how much to trust the result — unlike diff-in-diff, where a poor implicit match is easy to overlook.