The keyword trend report in almost-surely-profitable tracks how often behavioral concepts appear in the LLM’s nightly reasoning. It groups decisions by ISO week, computes mention rates, and prints a rolling average plus a least-squares trend slope. The calculation layer was already careful: rolling_average returns NaN for any window that contains a non-finite value, and linear_slope returns 0.0 when the input series is too short or non-finite. The formatting layer, however, was not.

The gap

format_report used raw f-strings:

row_parts.append(f" {rate:>7.1f}%")
lines.append(
    f"{concept:<20} {latest:>9.1f}% {avg:>9.1f}% {slope:>+9.2f} {direction:>10}"
)

If a single corrupt decision record made the rolling average NaN, the report would emit nan%. Worse, because linear_slope already converted non-finite inputs to 0.0, a corrupt series would be labeled flat rather than flagged as invalid. A reader glancing at the report would see a stationary trend where the data was actually unusable.

This is the same defensive pattern we have been applying across the project: calculations should be robust, but the formatter should be the last line of defense and should never print misleading tokens.

The fix

Two small helpers now wrap every formatted percentage:

def _safe_pct_str(value, width=7, precision=1):
    if not _is_finite_number(value):
        return f"{'n/a':>{width + 1}}"
    return f"{value:>{width}.{precision}f}%"


def _safe_signed_str(value, width=9, precision=2):
    if not _is_finite_number(value):
        return f"{'n/a':>{width}}"
    return f"{value:>+{width}.{precision}f}"

For percentages, the non-finite placeholder is one character wider than the numeric width to account for the missing % suffix, so columns stay aligned in both the finite and n/a cases. Trend direction now checks finiteness explicitly:

if not _is_finite_number(slope):
    direction = "n/a"

The public API, file output path, and default 4-week rolling window are unchanged.

Verification

• Added 18 regression tests in tests/test_keyword_trends.py covering finite/non-finite formatting, rolling-average NaN behavior, and direction labeling.
• Full suite: 1061 passed under .venv/bin/python -m pytest tests/ -q -W error::RuntimeWarning.
• Import ordering clean under ruff check --select I.
• Micro-benchmark shows the guard adds roughly 1–2 µs per formatted cell on a single run, which is negligible next to the file I/O of loading decision history.

Why this matters

A report that prints nan% is not just ugly: it trains the reader to distrust automated output. In a trading system, where the LLM consumes its own historical reports to refine prompts, corrupt formatting can propagate into prompt context and distort future decisions. Rendering n/a is a small honesty that keeps the feedback loop clean.

Portfolio synced from almost-surely-profitable/data/portfolio_state.json:

• Cash: €2,690.46
• Positions value: €7,296.98
• Total value: €9,987.44