The problem
Answering, weeks later, exactly why an automated system did what it did is harder than writing
its rules. If nobody can, people stop trusting it and start checking it by hand, and they're
back to the job they were trying to remove.
Today it's a research system on a demo account, which we use to study execution and risk
mechanics. It runs continuously on a hosted machine with live market data and simulated
execution.
The data
It reads live market data for one instrument and writes its own event log. We publish the
structure of the data and the log and nothing beyond it, and no source code: the rules
themselves are the part worth keeping private.
How we approached it
Every decision writes an event. The log has a locked schema, so a row written today can be
compared with one written six months ago. Adding a field takes a deliberate change; it can't
happen by accident.
Signals, rule evaluations and order events all land in the same place, in order. A rule that
fails is recorded just as carefully as one that passes, because the failures are what explain
the gaps.
The risk controls follow fixed rules, with no case-by-case judgment: volatility-scaled stops,
breakeven and partial-exit handling, and a health monitor that tracks recent closed trades and
sidelines a strategy that is behaving badly.
How it is built
It runs unattended on a hosted machine, and every strategy goes through one shared execution
loop. Each decision is written to an audit log with a fixed set of columns, so the behavior
can be reviewed afterwards without relying on anyone's memory.
Methods and tools: Python connected to MetaTrader 5,
volatility-based position and stop sizing, volume profile analysis, a regime filter, and
structured CSV audit logging with a fixed column schema.
What we can show
The view isolates one lifecycle: the signal, the pending order, the fill and the close, with
the rule evaluations recorded next to them and a timestamped table underneath. Any event can
be traced back to the conditions that produced it.
You won't find performance figures for this system here or anywhere else on this site. It runs
on a demo account, so there is no verified track record, and publishing simulated results as
though they were outcomes would be exactly the thing we criticize elsewhere.
What it demonstrates
Even if you have no interest in trading, the same idea applies to your business: every
decision the system makes is recorded, and any one of them can be reconstructed afterwards,
which is what makes an automated business process safe to leave running.
The logging and run guide in the automation package make
the same promise. This system keeps it at a harder standard, because its events arrive
continuously and can't be re-created by running the job again.