Why Statistical Process Control beats machine learning for field-force alerts
When a field manager in Mwanza gets an alert at 7am telling them an agent is at risk, they need to act on it immediately. But first, they need to trust it.
This is the fundamental problem with machine learning in field-force management: the alert comes from a model no one can explain. The manager asks: "Why is this agent flagged?" The system says: "Because the model says so." The manager stops trusting the alerts. The alerts get ignored.
The SPC alternative
Statistical Process Control was developed in manufacturing quality control in the 1920s. It is not new. It is not exciting. It is one of the most battle-tested statistical methods in the history of operations management — and it is exactly the right tool for field-force early warning.
The principle is simple: calculate a rolling mean for any metric over a defined period. Calculate the standard deviation. When a data point falls more than 2.3 standard deviations below the mean, raise a CRITICAL alert.
In TradeMetriq's implementation:
x̄ = 21-day rolling mean of daily visit completions
σ = standard deviation of that same period
CRITICAL threshold: x̄ − 2.3σ
When an agent's visit completion rate falls below that threshold, the system fires a CRITICAL alert. The manager can see the formula. The agent can see the formula. Anyone can verify the calculation with a calculator.
Why 2.3σ specifically?
At 2.3 standard deviations below the mean, a measurement is in the bottom 1% of the historical distribution. In field-force terms, this means the agent is performing at a level they have never reached in 21 days of operation. Something has changed. Something is wrong.
The threshold is not magic. It is a deliberate choice that balances two risks: missing a real problem (false negative) and generating so many alerts that managers stop acting on them (alert fatigue). 2.3σ hits the correct balance for the volume of agents and visit data typical in East African FMCG distribution.
The trust dividend
When managers can see the formula, they ask better questions. Instead of "why is this agent flagged," they ask "what changed in the last 21 days?" The alert becomes a conversation starter, not a black box output.
Field agents, when shown their own SPC-derived score, understand it immediately. "You've dropped below your 21-day average by more than normal" is a sentence any agent can process. It does not require a data science degree.
This is why TradeMetriq chose mathematics over machine learning. Not because ML cannot work. But because explainability is a feature, not a limitation — especially in East Africa, where trust between field teams and technology is still being established.