Research
The point of this page is finding the safest way to exit a losing position - balancing risk of ruin (holding a fading bucket too long) against risk of reward (cutting it before it recovers). First question: how far can a bucket drop in price from where it was picked before it becomes statistically unlikely to still win?
Four independent identification methods, reused directly from Market Knows - each picks one candidate bucket per city-day at the same 7am (10am-fallback) snapshot:
For every settled city-day where a method identifies a bucket, that bucket's own YES price is then tracked at every later snapshot the same day. Each later snapshot contributes one observation: how many points has the price fallen from its entry price so far (0 if it never fell), and did the bucket go on to win the day. Observations are binned into 10-point drop bands - the same city-day contributes many observations across the day, the same way the Timing heatmap pools multiple hourly snapshots per day, since the question here is inherently about what happens at every point in time a stop-loss line could have been crossed, not just one moment. Built from every Polymarket snapshot captured since 2026-05-22, through 2026-11-06.
X-axis is how far the picked bucket's own YES price has fallen from its entry price (10-point bands). Y-axis is how often that bucket still went on to win. One line per identification method - if a line falls off a cliff early, that method's picks rarely recover from even a small drop; if it stays higher for longer, that method's picks are more resilient and can tolerate a wider stop-loss.
Same observations as the chart above, but no longer pooled across every entry price - each method's own picks are cross-tabbed by the bucket's entry price (5-point bands, y-axis) against how far it has since dropped (10-point bands, x-axis). Checks whether a given drop means something different for a bucket that started expensive versus one that started cheap, the way Divergence's own price-controlled gap view does for gap size.
A stop-loss shouldn't be about waiting to see if a fading pick recovers - it's about telling a harmless dip apart from a real regime change. This cross-tabs the same drop bands (x-axis) against how far behind the market's own current #1 bucket the held pick has fallen (10-point bands, y-axis - 0 means it's still the market's own top pick, even after dropping from its own entry price). Computed at every later checkpoint regardless of whether the lead has actually changed hands, and pooled across all four methods rather than split - drop and leader-gap aren't obviously method-specific the way entry price was, and splitting four ways would thin out the far cells fast.
Price x Drop (per method): a given drop matters more for a cheap entry than an expensive one, but entry price mostly reshuffles when a drop starts to sting.
Drop x Divergence From Leader (pooled): this is the one that decides whether a drop is actually dangerous. A bucket barely down from entry but already 90+ points behind the market's current leader is already a near-certain loser (~0% win rate) - while a bucket still holding the market's own top spot tolerates a real drop reasonably well (67-73%).
Check leader-gap first. It's the closer proxy for "is this pick actually wrong" - drop is a secondary/severity signal, not an equal partner.
Follow-Up Question
Now that price drop and divergence-from-leader look like the right two conditions to gate a stop-loss on, the next question is what actually happens when they fire. Four options were on the table, but two of them are the same trade: buying NO on a position is economically identical to selling its YES (confirmed earlier - no_price ≡ 1 - yes_price in 99.9% of raw snapshots), so "sell current, buy YES on new winner" and "buy NO on current, buy YES on new winner" are one strategy said two ways. That leaves three genuinely distinct options:
Should the exit strategy be condition-specific? The evidence so far says no, at least not on which of the three to run - both attempts at "rotate" failed regardless of which entry condition they were attached to, which points at chasing an already-moved market being a bad idea in general, not a Confirmed-Bias-specific one. Condition-specificity looks like it belongs in where the threshold sits (Mkt/Sys Forecast tolerate deeper drops before their win rate collapses than Mkt/Sys + Bias do) rather than in which exit strategy fires.
Other exit strategies worth trying later:
Testing It
All three exit strategies gated on leader-gap, swept across the same candidate triggers, on the same pooled entries. Clean Close: at the first checkpoint where the held bucket falls that many points behind the market's current leader, sell its YES; otherwise hold to normal resolution. Rotate: same trigger, but instead of just selling, the freed capital immediately opens a new YES position on whichever bucket is currently leading, held to resolution with no further monitoring - one redeployed unit, not an extra one. Hedge: the original position is left completely alone (still held all the way to resolution), and a separate new NO position opens on the current leader at the same trigger - one more unit added, not a redeployment, so unlike the other two it's not capital-neutral against the baseline once it fires.
The dashed line in each chart is the shared baseline: no stop-loss at all, always holding the original pick to resolution. Clean Close is the untested other half of the earlier on/off comparison, now swept across a full range of triggers instead of one fixed price.
Loading sweep results...
Summary
Clean Close never beats simply holding to resolution, at any trigger tested. By the time a position has fallen far enough behind the leader to fire, its price has usually already cratered - selling mostly locks in a loss that's already largely realized, while holding on keeps the small remaining chance of the full payout that a close gives up for good.
Rotate is worse than Clean Close at every trigger, with a far deeper worst-case loss - closing out and immediately buying into the new favorite just compounds one losing leg with a second one, the same "chasing an already-moved market" failure this page's earlier stop-loss-rescue experiment and Confirmed Bias's own live rotation both already showed.
Hedge is the only one that beats doing nothing, at every trigger - but it isn't actually hedging anything. Checked against the real data at a 25pt trigger: the original position and the new NO-on-leader position win together 23% of the time and lose together 54.5% of the time - they're positively correlated in over three-quarters of outcomes, not offsetting. They only decorrelate (22.6%) when some third bucket wins instead of either one. Its edge comes from the ordinary bounded-downside/open-upside shape of a binary contract, stacked a second time on top of a pick that already had real statistical edge from the entry conditions underneath it - not from reducing risk. A more honest name for it would be something like "double down against the current favorite," not "Hedge."
Net result: nothing tested here functions as a genuine protective exit. The safest thing to do with a fading position, on this data, is still to hold it to resolution - the only strategy that beats that is one that adds a second, correlated bet rather than reducing exposure to the first.
Testing It
Flat Clean Close fires the same trigger regardless of how much of the day is left - one reason it never beat holding: by the time leader-gap is large enough to matter, the position has usually already cratered, so a flat trigger can't catch it any earlier just because it's late in the day. This variant starts loose (a 90pt leader-gap threshold - close to "don't cut early no matter what") at each city-day's first checkpoint after entry, and linearly tightens toward the swept late-day threshold shown on the x-axis by its last checkpoint - interpolated by position in that day's own checkpoint sequence, not clock hour, since a settled day's tail can spill past midnight. Everything else matches Clean Close exactly (sell the YES, no re-entry), so each point is directly comparable to Clean Close's flat sweep at the same threshold value.
Loading sweep results...
Testing It
Every sweep so far triggers on leader-gap. This one switches to price drop instead, to isolate one specific question: does anchoring the drop measurement to the position's own PEAK price since entry - not the static entry price - catch a bucket that ran up first and then faded earlier than the current entry-anchored measurement would? Entry-Anchored: drop = entry price minus live price, same metric this page's very first descriptive chart used, now run as an active policy for the first time. Can never trigger while a bucket is still above its own entry price. Trailing: drop = the running peak price since entry (only ever ratchets up as new highs appear) minus live price - can trigger even on a bucket still above its own entry price, as long as it's fallen this many points from its own high. Same Clean Close exit mechanics (sell the YES, no re-entry) for both.
Loading sweep results...