News
A dated record of what actually reached production, newest first, including the experiments that did not work. Items marked "in progress" carry no date on purpose: we do not promise dates we cannot keep. Everything below is drawn from the platform's own repositories and deployment history.
Before a rebalance, the execution service will reconcile its own record of working orders with what the broker reports, and cancel stale orders instead of trading around them. The change is written and under review; it ships when the review is done, and no date is promised.
Any orchestrated flow that fails, crashes or times out posts to the monitoring channel automatically. Until now a failed morning run could go unnoticed until someone looked at the dashboard; now the alert arrives before the market opens.
For two weeks in August the morning feature build occasionally overlapped with inference, inference saw a partial cross-section, and the book correctly fell back to cash. The technical features are now computed the evening before, and every downstream step is chained on the completion of the one before it rather than on a clock.
The balanced long and short book became the only portfolio the platform trades. Shorting sits behind an explicit switch, and a book containing a side the platform is not cleared to trade stops the whole run rather than trading half of it.
We built news sentiment end to end and tested two independent ways of reading the news, then trained a candidate model on the result. Neither version improved on the model already in production. We removed the signal, wrote down the finding, and left the groundwork in place in case a longer horizon changes the answer.
The platform began executing every trading day, tracking the target book position by position. Account value is read in the account's own currency before every run, and the performance baseline is anchored per account epoch so a reset can never appear as a phantom gain or loss.
The founders now have a single view of the daily cycle: what ran, what failed, and the work each day's errors turned into. The dashboards are defined as code alongside everything else, so what is on screen always matches what is in version control.
With the brokerage connection running on Google Cloud, the platform read the live account, priced the target portfolio and computed a full set of orders for the first time. The gross-exposure circuit breaker correctly stopped the run, which is exactly what a first run should do, and surfaced a sizing question that was resolved before any order was placed.
The whole platform moved to Google Cloud in a European region, with the databases migrated intact. The cloud environment is declared as code and applied automatically after review, and continuous integration authenticates without any long-lived keys.
A new stage was added between the forecasts and the broker: it turns scores into signals, signals into target weights within firm limits, and applies the risk gates before publishing the day's book. It decides what to hold; the execution stage decides how to get there.
Work began on the stage that connects the platform to the broker. It started as an order model and a service skeleton, and became the rebalancer that trades the portfolio today.
The pieces every stage needs, from database access to logging to the checks that gate a release, moved into one shared foundation that carries no trading logic of its own.
Hieronim Kubica and Michał Szarek founded Curson Capital in Poland to build and run a systematic trading platform end to end, and registered cursoncapital.com. The company is bootstrapped and has taken no outside investment.
The stage that trains and serves the machine-learning models began as a training pipeline with every experiment tracked. The rule that decides which model trades, and when a candidate replaces it, came later.
The first two stages, market data and signals, were created on the same day. Everything the platform does today rests on the daily prices and the signal families these two have computed since.
Want to hear about the next entry? Email office@cursoncapital.com. We do not run a newsletter or a mailing list; a reply is the whole system.