Position sizing and stop placement
from MAE/MFE distributions.

Why a return distribution alone isn't enough for sizing

The single-symbol use cases built on Konseki data — chat answers, pre-trade context, signal explanations — mostly draw on the forward return distribution: best case, worst case, percentile spread, positive rate. That's the right data for answering "what historically happened next." It's the wrong data, on its own, for answering "how much should I risk, and where should my stop go":

Two setups can have an identical positive return rate and an identical median return while having completely different paths to that outcome. One might have moved steadily in the favorable direction the whole way. The other might have moved sharply against the position first, then recovered. A trader using a stop placed without accounting for that difference would have been stopped out of the second setup before it ever reached the outcome the return distribution is describing.

What MAE and MFE actually measure

These are path statistics, not endpoint statistics:

returns.max_adverse_excursion (MAE) reports the worst point a historical match moved against the position before the forward window closed — the deepest unfavorable excursion along the path, not the final outcome. returns.max_favourable_excursion (MFE) reports the opposite: the best point the position reached in its favor during the same window, again independent of where it ended up by the close of the period.

A setup with a strong positive median return can still carry a meaningfully negative MAE, meaning that even the historically favorable instances of this setup often moved against the position first. Reading return distribution alone would miss this entirely, because the return distribution only reports where each instance ended up, not where it went on the way there.

Using MAE to reason about stop placement

A stop placed tighter than the historical MAE distribution for a given setup risks being triggered by normal, expected adverse movement — not by the setup actually failing. If the worst historical instances in the match set moved against the position by some meaningful percentage before resolving favorably, a stop placed inside that range would have closed out a position that was still on track, based on what the historical evidence shows for setups structurally similar to this one.

This doesn't mean a stop should always be placed at or beyond the worst historical MAE — that depends on risk tolerance and how much weight a given match set deserves as evidence in the first place, which is where match_quality and evidence_count come back into the picture:

It does mean a stop set without reference to the MAE distribution is a guess, where one set with reference to it is grounded in what comparable historical setups actually required to play out.

Using MFE alongside MAE for sizing, not just exits

MFE matters for position sizing in a less obvious way:

Comparing MAE and MFE together gives a rough sense of the risk-to-reward shape of a setup before any position is taken: a setup with a shallow typical MAE and a generous typical MFE is offering a different risk profile than one where both are similarly sized, even if their final return distributions look comparable. Sizing decisions made purely from the return distribution miss this shape entirely.

Why this needs the full match set, not a single data point

MAE and MFE are most useful read as distributions across the matched historical instances, not as single numbers. The reported max_adverse_excursion reflects the worst case across the match set — useful as a boundary, but a sizing decision benefits from understanding how typical that worst case actually was, which requires looking at the spread across matches rather than the single extreme value alone.

What this looks like in practice

A position-sizing tool built on Konseki data pulls the MAE and MFE figures for the relevant forward window alongside evidence_count, diversity.score, and match_quality for the matches feeding that window. A stop is set with reference to the historical MAE distribution rather than an arbitrary percentage. Position size is scaled based on how well-supported that MAE figure is — a wide stop justified by a deep, diverse match set is a different decision than the same wide stop justified by a handful of thin matches, even though the dollar risk on the position might be identical in both cases.