Speed is a capability

Many people who work in Southeast Asia notice one outstanding trait of Chinese teams: they are fast. They decide fast, launch fast and adjust fast. If a campaign does poorly today, the creative is replaced tomorrow; if a channel brings no volume this week, next week there is a different channel. Local colleagues are sometimes impressed and sometimes exhausted.

Why is speed an advantage, and when does it become a cost? Control theory provides a clear framework: treat a team as a feedback control loop. It observes the market (the sensor), compares what it sees with its goal (the error), adjusts its practice by some amount (the gain), and observes again. Two parameters of the loop are crucial (Åström and Murray 2008):

• Delay $\tau$: how long it takes from a change in the market until the team’s adjustment actually takes effect, including reporting cycles, layers of approval and preparation for execution.
• Gain $k$: how large an adjustment the team makes each time it sees a gap. A gain of 1 closes the whole visible gap at once; a gain of 0.2 closes a fifth of it each time.

Tracking a moving target

In a fast-changing market, the best practice (price, product mix, lead message) itself drifts. The team’s job is not to find an optimum and stop, but to keep tracking a moving optimum.

# Iteration speed as a control problem. The best offer for a market (price point, product mix,
# message) drifts like a random walk (step SD 1 per week). The team sees the market with noise
# (SD 2) and corrects its position by gain k times the gap it saw, but decisions act on data that
# are tau weeks old (reporting cycle + approval chain). Root-mean-square tracking error over
# 4,000 weeks, averaged over 20 runs; 'unstable' means the error grew without bound.
import math, random

def rms(k, tau, weeks=4000, runs=20):
    tot = 0.0
    for seed in range(runs):
        rng = random.Random(seed)
        target = 0.0; xs = [0.0] * (tau + 1); seen = [0.0] * (tau + 1); err2 = 0.0
        for t in range(weeks):
            target += rng.gauss(0, 1)
            seen.append(target + rng.gauss(0, 2))
            x = xs[-1] + k * (seen[-1 - tau] - xs[-1 - tau])
            xs.append(x)
            if abs(x - target) > 1e6: return float("inf")
            if t > 200: err2 += (x - target) ** 2
        tot += math.sqrt(err2 / (weeks - 201))
    return tot / runs

print("delay (weeks) |  gain 0.1  gain 0.2  gain 0.4  gain 0.6  gain 0.9  gain 1.2")
for tau in (0, 1, 2, 4, 8):
    cells = []
    for k in (0.1, 0.2, 0.4, 0.6, 0.9, 1.2):
        e = rms(k, tau)
        cells.append("unstable" if e == float("inf") or e > 200 else f"{e:8.1f}")
    print(f"{tau:13d} | " + "  ".join(f"{c:>8s}" for c in cells))
# Stability limit for x_{t+1} = x_t - k x_{t-tau}: k < 2 sin(pi / (2 (2 tau + 1)))
print("largest stable gain:", ", ".join(f"tau {tau}: {2 * math.sin(math.pi / (2 * (2 * tau + 1))):.2f}"
                                       for tau in (0, 1, 2, 4, 8)))
# delay (weeks) |  gain 0.1  gain 0.2  gain 0.4  gain 0.6  gain 0.9  gain 1.2
#             0 |      2.1       1.5       1.2       1.4       1.8       2.5
#             1 |      2.2       1.7       1.7       2.3       5.4  unstable
#             2 |      2.4       2.0       2.5       9.2  unstable  unstable
#             4 |      2.7       2.7  unstable  unstable  unstable  unstable
#             8 |      3.6  unstable  unstable  unstable  unstable  unstable
# largest stable gain: tau 0: 2.00, tau 1: 1.00, tau 2: 0.62, tau 4: 0.35, tau 8: 0.18

The table shows three things.

First, shortening delay has value. At the same gain of 0.2, raising the delay from 0 to 4 weeks raises the tracking error from 1.5 to 2.7; at a delay of 8 weeks a gain of 0.2 already exceeds the stability limit, and the team falls into ever larger swings. In this sense Chinese teams’ habit of compressing approval chains and shortening reporting cycles really does improve tracking.

Second, speed is not only short delay but also the right gain. With no delay, the best gain is about 0.4 (error 1.2); at a gain of 1.2 the error doubles (2.5), because the team is chasing noise in the data.

Third, and most important: delay and gain must match. With a delay of 1 week the gain must not exceed 1; with 2 weeks, 0.62; with 4 weeks, only 0.35. A team used to short chains and high gains at home that enters a cross-border structure with a longer chain (weekly local reports, fortnightly regional meetings, headquarters approval) and keeps the same correction strength will show the classic symptom of oscillation: scale up last month, cut this month, scale up again next month. Part of what local partners see as erratic decision-making may be exactly this instability of the control loop, not anyone’s character.

The cost of turning: brand assets

The model above assumes that changing practice is itself free. In many markets it is not. Customers’ knowledge of a brand builds up over time; Keller calls this accumulation customer-based brand equity, arising from the awareness of and associations with the brand held in customers’ memory (Keller 1993). Every visible turn (new positioning, new lead message, new image) may partly erase what has accumulated.

# When fast turning costs more than it gains: positioning changes erode accumulated brand assets.
# The best positioning drifts slowly (random walk, step SD 0.3 per month); monthly sales data show it
# with noise (SD 2). Revenue each month = (base + brand stock) * fit, where fit = exp(-gap^2 / 8)
# for the gap between the firm's positioning and the true optimum. Brand stock grows by 1 a month
# and keeps 97 % of itself, but each month's repositioning |dx| destroys a share min(1, w * |dx|).
# w measures how much the market rewards consistency: near 0 for pure performance-marketing
# categories, larger where customers buy reputation (premium, trust-heavy or relationship goods).
# Policies: chase (jump to last month's noisy reading); smooth 20 % or 5 % (move that share of the
# way: an exponential filter); step (keep a 10 % filtered estimate in the background and reposition
# to it only when it has moved more than 1.5 away - change rarely, in planned steps); freeze (never
# move). 600 months x 30 runs.
import math, random

def run(policy, w, months=600, runs=30):
    total = 0.0
    for seed in range(runs):
        rng = random.Random(seed)
        opt = x = brand = est = 0.0
        for t in range(months):
            opt += rng.gauss(0, 0.3)
            seen = opt + rng.gauss(0, 2)
            est += 0.1 * (seen - est)
            new = {"chase": seen, "smooth 20%": x + 0.2 * (seen - x), "smooth 5%": x + 0.05 * (seen - x),
                   "step": est if abs(est - x) > 1.5 else x, "freeze": x}[policy]
            brand = (0.97 * brand + 1.0) * (1 - min(1.0, w * abs(new - x)))
            x = new
            total += (5 + brand) * math.exp(-(x - opt) ** 2 / 8)
    return total / (months * runs)

POL = ("chase", "smooth 20%", "smooth 5%", "step", "freeze")
print("value of consistency w |  " + "  ".join(f"{p:>10s}" for p in POL) + "   (mean monthly revenue)")
for w in (0.0, 0.05, 0.2, 0.5):
    print(f"{w:22.2f} |  " + "  ".join(f"{run(p, w):10.1f}" for p in POL))
# value of consistency w |       chase  smooth 20%   smooth 5%        step      freeze   (mean monthly revenue)
#                   0.00 |        25.9        34.1        32.8        32.0        13.9
#                   0.05 |         8.1        23.6        29.3        30.5        13.9
#                   0.20 |         4.6        13.5        22.6        26.6        13.9
#                   0.50 |         3.9         8.6        16.0        21.1        13.9

When brand assets do not matter ($w = 0$), smoothing at 20 % is best; chasing the data is worse, because the data are too noisy. The more brand assets matter, the more the best policy leans towards adjusting rarely and in steps: at $w = 0.2$ the step policy earns 26.6, almost six times the chasing policy’s 4.6. Freezing is never best: a firm that never adjusts is eventually left behind by the market’s drift.

This adds an important condition to “fast iteration”: in markets where brand assets must be built over the long term, the best speed is fast behind the scenes and steady in front of customers. Data can be collected and analysed daily, but repositioning visible to customers should be rare and planned, and preferably agreed in advance with partners as to when it will happen, which brings us back to the conclusions of Chapters 2 and 3.

Transaction costs: why not everything should be negotiated

Williamson’s transaction-cost economics explains the same phenomenon from another angle (Williamson 1985). Whether a transaction is best organised through the market, a contract or a hierarchy inside a firm depends on three features: asset specificity (how much an investment made for this transaction is worth outside it), uncertainty and frequency. The higher the specificity, the greater the uncertainty and the more frequent the transaction, the more it pays to use tighter governance (long-term contracts, joint ventures, even integration) to protect each side’s investments and to reduce the cost of renegotiating each time.

Seen this way, the trouble with speed in cross-border ventures is this: if every adjustment requires renegotiation (another meeting, another confirmation of responsibilities, another explanation), the transaction costs of frequent adjustment pile up fast and use up the partner’s patience. In transaction-cost language, the written milestones and residual-control arrangements of Chapter 3 turn frequent adjustments from negotiation into execution: agree on the rules once, and afterwards each adjustment simply follows the rules without being renegotiated.

Summary

Control theory splits speed into two quantities: delay and gain. Shortening delay almost always helps; the gain must match the delay, or a team will oscillate in a cross-border structure. In markets where brand assets must be built, the best rhythm is fast behind the scenes and steady in front of customers. Transaction-cost economics adds that agreeing the rules of adjustment in advance keeps the cost of frequent adjustment to a minimum. The next chapter turns to a wider question: how great is the institutional distance among the countries of Southeast Asia, and between them and China?