Prediction Track Record

• Graded as published: a call is scored as it went out. The call itself is never edited, withdrawn, or removed once published.
• Grading bugs are fixed rather than left in place, and every fix is listed under Corrections below, with its date and what it did to the total.
• Entry and settlement: each play settles on its own entry session.
• Voided calls: when a company reported earnings outside its holding window but within 21 days of it, the call is voided and counts as neither a win nor a loss. Price drift alone cannot settle a call whose catalyst never arrived. A report a whole quarter away says nothing about that window, so those calls settle normally.
• What else is absent, and why: a call is not counted while it waits on its expiry. Past expiry it is scored, unless something needed to settle it is missing, such as the published strike or the closing price, in which case it stays out rather than being guessed at. The weekly analyses also record names we looked at and passed on, and a decision not to trade is not scored as a call.
• P&L basis: spread P&L is modeled from real closes around expiration. Win or Loss is the sign of that modeled P&L.
• Where the modeled credit is weakest: the credit assumed for a spread is a fixed fraction of its width, calibrated from credits quoted in our own notes at the strikes those notes placed. When a strike is re-derived to a different distance from the price, that assumption no longer fits it, and a strike moved further out of the money is credited more than it would really have earned. QCOM in the week of 28 July is the clearest case: sound as a win, because the option expired worthless, but its modeled dollar figure is generous.
• Losses included: every settled call appears, with sample sizes shown.
• Scope: this ledger counts published SeerForecast calls. The in-app Model Learning panel counts every lane the model trains on, including the retired OpenAI engine's frozen history, so its totals are larger and its win rate differs from this page's.
• Benchmark: the equity curve overlays SPY over the same period.
• Audit it yourself: the full ledger is downloadable as CSV.
• Archived off-site: we file copies of this page with the Internet Archive (opens in a new tab) after each week's settlement, and the calendar there shows every capture. We cannot alter a stored copy, and we will not ask for one to be removed, so the promise above that a published call is never edited is one you can check against a record we do not host. What moves between two captures is the grading, never the call, and every correction is listed below.

Corrections

On 1 August 2026 we fixed three bugs in how the ledger read our own published notes. None of them changed a call. All three changed how a call was scored. Together they moved the published total from $12,277 to $12,460, a rise of $184, and the settled record from 80 wins and 27 losses to 79 and 27.

• Strike prices were being read out of dates. An entry line reading "Mon 7/6 to Tue 7/7" was parsed as a $7 strike, and a $62 stock cannot close below a $7 barrier, so the play scored a win it could not have lost. Three calls were priced this way. Regrading them against their real strikes lowered the total by $939, and turned one published win into a loss.
• One call carried another play's structure. A parser fault copied a neighbouring alternate onto a halo trade, which was then scored as a spread it never was. Read correctly it is a long call, and long calls carry no modeled P&L, so it now shows an outcome with no dollar figure. This lowered the total by $188.
• Strikes were read from the stale half of the note. Each note states placement twice, as a percentage below or above spot and as an approximate dollar strike worked out on the Friday before. The dollar figure goes stale when the stock moves before entry. On QCOM the reference was 7% high by Monday, which left a strike the note placed well clear of the price sitting on top of it. Grading against the stated percentage instead raised the total by $1,311, and turned one published loss into a win.

The last of those three works in our favour, which is why it is written out here in the same detail as the two that did not.

On 18 August 2026 we fixed a fourth, which moved no total on this page but did move what three rows show. Three calls from the week of 25 May were dated to a session the market never opened. A call on a company reporting before the open is entered the session before the print, and the code that worked out that session skipped weekends but not holidays, so all three landed on Memorial Day. The price lookup then resolved forward to the next available close. For a before-open reporter that is the session which already contains the earnings move, and the strikes are worked out as a percentage of the entry price, so they moved with it. AZO shows the size of it: entered at $3,100.11, the close after a 9% fall on its own print, when the last real session before it closed at $3,406.50.

Regrading the three against the correct session changed no outcome and no dollar figure, so it left the settled record and the published total exactly where they were. All three were wins before and after. What changed is the entry date and the strikes shown against those rows. Each of the three closed far beyond its winning strike in both placements, and a call that finishes clear of its strike books the same modeled result wherever inside that zone the strike sat, so moving the strikes moved no dollars.