Signal Postmortem compares the direction a signal predicted with the return that followed after 5 and 20 days. Each signal is classified as a true positive, false positive, missed opportunity or regime mismatch, and the record keeps the market regime at signal and exit.
Aggregated results can be turned into weight adjustment suggestions for edge-signal-aggregator and into skill improvement backlog entries. Exit prices come from the FMP historical price API when a key is set, or from values you enter by hand. The output is a review of past signals and not investment advice.
What You Get
- One JSON postmortem record per signal with realized 5d and 20d returns and an outcome category
- Summary statistics grouped by skill, ticker or month in a Markdown report
- Weight adjustment suggestions for edge-signal-aggregator, held back until at least 20 signals exist
- Skill improvement backlog entries in YAML for patterns such as false positive clusters
Common Use Cases
- Reviewing a batch of signals that have reached their holding period
- Recording a manual exit for one closed trade from a saved signal file
- Finding which source skill produces the most false positives in a given regime
What Makes It Unique
It separates a failed signal from a regime change by storing the regime at signal and exit. It also skips and counts any signal that lacks observed entry and exit prices, so a missing price never turns into a neutral outcome.
Under The Hood
Python scripts read signal JSON files from a directory and write postmortem JSON, Markdown and YAML files to a reports folder. With an FMP API key taken from the environment or a command-line option, the recorder calls the FMP historical price endpoint over the network using the requests library; without a key it works offline from manually supplied exit prices. The SKILL.md also tells the model to run these scripts from the shell.