A longer Tokyo opening auction cuts volatility over the first 30 minutes by about 21%, relative to market-wide volatility over the same minutes. Kasahara and Yamada's comparison accounts for the later start. Their 27% headline overstates what I would carry onto a trading desk, though the adjusted effect remains large.
Selection has always complicated the debate over whether trading pauses merely defer volatility. Pauses trigger when volatility is already high. Naive comparisons therefore find more volatility after pauses, as earlier U.S. research did. Hautsch and Horvath (2019), for example, constructed Nasdaq controls using periods before pause rules took effect. Kasahara and Yamada draw their counterfactual from within the same rules, using a within-regime IV and proprietary TSE order-book data covering 2014 to 2021.
Most pauses extend the opening auction
The Tokyo Stock Exchange (TSE) signals a single-stock trading pause with a displayed bid or offer, the special quote. The trigger is a prospective trade outside a range of one range step L around a reference price. Price brackets determine L, a fixed yen amount: 30 yen for stocks between 1,000 and 1,499 yen, and 40 yen from 1,500 to 1,999. Orders can still enter the book during the pause. Every three minutes, the exchange widens the range until the auction can clear.
The authors study a top-500 sample of First Section stocks, the exchange's main market segment until April 2022, plus J-REITs. It covers January 2014 to March 2021: 1,763 trading days and about 873,000 stock-dates. Fully 86% of pauses occur at the 9:00 open, before any trading in the stock. Those opening pauses number 77,528, or 8.9% of stock-dates, and their median duration is 3.0 minutes.
A "pause" in this sample usually means a few extra minutes in the opening call auction. That distinction governs how to read the results. The paper asks whether extending the call auction improves the open, a question closer to Madhavan's batch-auction argument than to the circuit-breaker debate.
The yen schedule supplies the identification. DistanceToSQ, the authors' instrument, divides L by the reference price and measures how far away the trigger is. At 1,490 yen, a stock's trigger is about 2.0% away; at 1,510 yen, it is about 2.6% away. Nearby prices thus face different pause probabilities, determined before orders arrive.
With firm and date fixed effects, the IV estimate for log 30-minute realized volatility is −0.311 (s.e. 0.051). Its Anderson-Rubin set is [−0.41, −0.21], and the first-stage F is 672.3. OLS gives +0.153 under the same specification. That reversal is the paper's strongest result: selection explains most of the naive correlation between pauses and volatility. The IV estimate also puts the decline in quoted spreads at 6.249 bp, against a sample mean of 16.1 bp.
What survives the later start?
The 27% volatility reduction needs a clock adjustment. Paused stocks begin trading later, when market-wide volatility is generally lower. The authors make that comparison themselves: relative to market-wide volatility over the same minutes, the coefficient is −0.230 (s.e. 0.051), or about 21%.
I would use 21%.
The effect then fades. For a window beginning at the open, the coefficient is −0.31; for one beginning 90 minutes later, it is about −0.07. The confidence band remains below zero through 75 minutes. Next-day realized volatility also falls (−0.14, t = −2.8). That finding weighs against the claim that pauses simply delay the move.
These are local estimates. The compliers are stock-dates whose pre-open order imbalance implies a move close to the range step, roughly 1.5% to 2.7% of price. The authors explicitly leave effects during extreme shocks unestablished. Price position also changes pause probabilities only slightly: the instrument's partial R² is 0.010.
Measurement creates another reservation, one the authors identify as their main limitation. After a pause, the share of minutes without a midpoint change rises (IV 0.39, t = 5.6), chiefly among coarse-tick stocks. Transaction-price volatility nevertheless falls by 0.253 log points.
The decomposition suggests fewer round trips. Log path variation falls 0.575 (t = −5.7), whereas log squared net return falls 0.212 (t = −1.6, insignificant). The path zigzags less, with no statistically distinguishable change in its destination. As the authors caution, this decomposition does not separately identify permanent and transitory components.
Liquidity waiting beyond the special quote
The order-level results fit the explanation that extra auction time helps. During the first three-minute cycle, 77.9% of new limit orders opposing the imbalance are priced more than half a range step beyond the special quote. Within that 77.9%, orders already beyond the quote's position after the next widening account for 48.2% of all such orders.
Order arrivals in later cycles are nearly uniform. The share arriving within 30 seconds after an update is 18.8%; within 30 seconds before the next update, it is 18.5%. An even distribution would give 16.7%. Successive widenings move the quote toward liquidity already in the book.
Before the stock opens, about 14% of the 9:00 matching price move reverses (slope 0.859, R² 0.936). The IV efficiency test finds a worse 9:00 matching price for paused stocks (interaction −0.202), but no detectable gap at the actual opening price (−0.020). The authors treat the order-flow evidence and within-pause comparisons as descriptive. That is the right limit to put on them.
Execution desks should pay more attention to the differences across events than to the average. Following overnight earnings news, an interaction of +0.328 offsets a main effect of −0.317, leaving roughly nothing. A negative own overnight return reduces the benefit by 0.165. Large market moves reduce it by 0.254.
This pattern fits extensions doing most good when the imbalance reflects a temporary shortage of offsetting orders. After earnings, they do roughly nothing; after large market moves, the effect falls to about −0.15 log points. For TOPIX 100 names, the 100 largest, most liquid TSE names, volatility still declines (−0.241). Quotes do not get cheaper: the spread interaction of +13.654 bp against a main effect of −8.998 bp leaves no narrowing. Large-cap desks get a quieter price path without tighter spreads. Traded value is insignificant (0.166, s.e. 0.139), though the authors stress that economically meaningful declines in activity remain possible.
Mid-session evidence is too imprecise
Pauses during continuous trading interrupt an active market, and the evidence here is weak. A price-driven pause occurs when an incoming order would trade outside the range. Across 1,406 events, its estimated effect is 0.018, with an AR set of [−0.52, +0.48].
Order-book-evaporation pauses occur when the opposing side has emptied, leaving no matching price as the special quote appears. Their estimate is −0.577, with [−1.72, −0.01], across 2,032 events. The first-stage F statistics are 30.1 and 19.6. Both sets include the morning-open estimate, as the authors acknowledge, so the analysis establishes no difference across sessions. Only 56 such events fall within the top 500, forcing this part of the study into smaller stocks.
The primary exchange's opening auction is exempt from the U.S. Limit Up-Limit Down (LULD) rule. The authors accordingly write that their opening-auction estimates "should not be read as evidence on LULD pauses."
We could not test any of these results ourselves. Doing so requires TSE tick-level order books with special-quote flags and the hypothetical matching price during each pause. Minute bars cannot reconstruct an auction before it prints.
The paper reports neither trading-strategy returns nor transaction costs. It offers a prior for sizing 9:00 Tokyo risk. When a special quote accompanies a moderate imbalance, expect volatility in the first half hour to be about a fifth lower than without the extension, relative to market-wide volatility over the same minutes. The open will still be noisier than a normal one, as the positive OLS coefficient (+0.153) indicates.
Evidence after March 2021 or under a different update interval is the one result that would change my mind. The design keeps the three-minute schedule fixed. It cannot establish whether the auction clears at the right moment.