Building a local investment research project
This home-server project brings together portfolio records, model-generated research and simulated investment pots. The public pages show dated observations from the experiment. They do not establish a profitable strategy.
The main parts
- A portfolio tracker records holdings.
- Local and API language models produce research notes and briefings.
- Separate strategy rules manage paper trades using virtual money.
Keeping these parts separate makes it easier to distinguish a recorded fact, a model’s interpretation and an action taken by a strategy.
Record actions and failures
The system records an intended order before attempting it, then reconciles the result. If the data or journal cannot be read reliably, a trading path should stop rather than infer that it is safe to continue.
The build notes record a failed order caused by a non-finite price. Recording the attempt and outcome made it possible to trace the problem. Missing market inputs also need an explicit unavailable state, rather than a confident-looking summary.
Keep unproven signals out of trading decisions
One off-exchange-volume signal was replayed against historical data. The recorded result did not establish statistical evidence for an edge, so the signal remained context-only. A backtest and a short forward record are limited evidence; neither guarantees future results.
Review the code
AI tools helped write and review code. A review identified problems involving spending limits, paths and external data displayed in the dashboard. Fixes need tests and verification: a second model can share the first model’s mistakes.
Run the workloads within the server’s capacity
The local models share a 32 GB GPU. A memory-management proxy queues large requests and releases eligible idle models. Longer research tasks may use external APIs, which adds a separate cost and data-processing route.
What remains to be measured
Electricity, hardware, API use and maintenance all contribute to operating cost. Those costs have not been fully measured, and the public experiment has not established that its strategies are profitable.
You can explore the virtual investment pots and their recorded values. They are an engineering experiment using virtual money, not a recommendation to invest.