Scoreboard

AI news forecasts: the full record

Before each session an AI model reads the news and forecasts whether EUR/USD, EUR/CHF and USD/CHF will go up or down. Every forecast is listed here once its window has closed, hits and misses alike, with the reasoning behind it and what actually moved the price. The method has been revised several times; the numbers below combine all versions, and each version is also scored on its own. Background: Can AI predict EUR/USD from the news? 120 forecasts later

AI news forecasts · EUR/USD, EUR/CHF, USD/CHF
coin 50%
50.3%
hit rate · 86 of 171 forecasts · all method versions
50.3%AI
50%coin flip
43%–58%95% range
95% range: the true hit rate is very likely inside it
Can't be told apart from guessing — yet

Hit rate as forecasts add up

shaded = 95% range · dashed = coin flip
5 Oct · forecast 1 16 Nov · forecast 171

Early on a few forecasts swing the rate wildly; the range narrows as more come in. A range that still includes 50% means the result can't be told apart from guessing.

By pair

bar = hit rate · whiskers = 95% range · dashed = coin flip

EUR/CHF 51%
EUR/USD 54%
USD/CHF 46%

By session

EU 08–12 UTC · US 12–21 UTC · Day = 08–21 UTC, used only 5–8 Oct

US 51%
EU 54%
Day 25%

Misses: how sure the AI is of the reason

after each window the AI explains the move and rates how sure it is of that explanation

Clear reason: 7Plausible reason: 35Unclear: 4385misses
  • Clear reason 7
  • Plausible reason 35
  • Unclear 43

Day by day

one cell per forecast · EU first, then US · hover a cell for details hit miss
5 Oct
6 Oct
7 Oct
8 Oct
9 Oct
12 Oct
13 Oct
14 Oct
15 Oct
16 Oct
19 Oct
20 Oct
21 Oct
22 Oct
23 Oct
26 Oct
27 Oct
28 Oct
29 Oct
30 Oct
2 Nov
3 Nov
4 Nov
5 Nov
6 Nov
9 Nov
10 Nov
11 Nov
12 Nov
13 Nov
16 Nov

Versions and weekly audits

now on v6

The forecasting method has a version number, shown on every call below. Every Saturday an AI agent audits the past week: it looks at what went wrong, tests new ideas on years of price history, and has a separate skeptic agent try to break each result. A new version ships only when an idea passes both; most weeks nothing does. Versions are scored separately and never mixed, so a change can't hide in an average. We audit the method; we don't promise it gets better. How the weekly audit works.

VersionIn useCallsRightWhat changed
v6 2 Nov – 16 Nov 63
31 (49%)
Tested 4 ideas; one passed: on US federal holidays the US session is much quieter, so its expected size is scaled down. Direction calls are unchanged. Rejected: a direction-aware edge flag, prices reversing after central-bank decisions, and high currency volatility predicting continuation.
v5 26 Oct – 30 Oct 30
16 (53%)
too few calls to compare
The first weekly audit, together with six studies run the same week. Version 5 adds an expected size of each session, moves in bonds, stocks, oil and gold as context, and a lower confidence cap on days the Fed chair speaks. One tested idea became a rule: statistical edge flags no longer tilt a call, because the skeptic showed they only echoed each pair's long-run drift. Three ideas were rejected.
v4 15 Oct – 23 Oct 42
17 (40%)
too few calls to compare
The forecaster reads the day's economic calendar and has to search properly before calling. On two days it had made every call after one stale search, including on a US inflation release day.
v3 12 Oct – 14 Oct 18
14 (78%)
too few calls to compare
A driver such as bond yields can back a call only through how it moved since the previous close, never through its level alone.
v2 8 Oct – 9 Oct 9
6 (67%)
too few calls to compare
Each call also gets statistics from 15 years of hourly prices, such as how often the pair rose in this window after a similar move.
v1 5 Oct – 7 Oct 9
2 (22%)
too few calls to compare
The first version: before each session the AI reads the morning news and calls each pair up or down, with no price statistics.

Below about 50 calls a version's hit rate is mostly luck: the 95% range is wider than ±14 points, so two versions ten points apart can't be told apart. The numbers are shown so nothing is hidden, not as proof that one version is better.

Change log

DateWhatOutcomeDetails
15 Nov 2026
Weekly audit 2026-W46
3 ideas tested · 0 passed
no new version Tested 3 ideas; none passed, so the method stays at version 6. Rejected: the reaction to a central-bank statement reversing later in the session, bond-market uncertainty making decision days bigger, and the first hour after a Swiss rate decision setting the afternoon direction.
7 Nov 2026
Weekly audit 2026-W45
4 ideas tested · 0 passed
no new version Tested 4 ideas; none passed, so the method stays at version 6. Rejected: a run of US data surprises setting the dollar's direction, a late move before the European open fading, and Swiss bank deposit data predicting the franc. A volatility idea looked promising but is now checked on new data before it can change any call.
31 Oct 2026
Weekly audit 2026-W44
4 ideas tested · 1 passed
v6 shipped Tested 4 ideas; one passed: on US federal holidays the US session is much quieter, so its expected size is scaled down. Direction calls are unchanged. Rejected: a direction-aware edge flag, prices reversing after central-bank decisions, and high currency volatility predicting continuation.
26 Oct 2026
Lessons v2
approved by a person
lessons v2 on The first lesson was reworded so it no longer depends on statistical edge flags, which stopped tilting calls in version 5.
24 Oct 2026
Weekly audit 2026-W43
4 ideas tested · 1 passed
v5 shipped The first weekly audit, together with six studies run the same week. Version 5 adds an expected size of each session, moves in bonds, stocks, oil and gold as context, and a lower confidence cap on days the Fed chair speaks. One tested idea became a rule: statistical edge flags no longer tilt a call, because the skeptic showed they only echoed each pair's long-run drift. Three ideas were rejected.
22 Oct 2026
Lessons v1
approved by a person
lessons v1 on First lesson switched on: when there is no dated news for the window, the call stays at 50% and says it is a coin flip.
15 Oct 2026
Method v4
by hand, before the weekly audits
v4 started The forecaster reads the day's economic calendar and has to search properly before calling. On two days it had made every call after one stale search, including on a US inflation release day.
12 Oct 2026
Method v3
by hand, before the weekly audits
v3 started A driver such as bond yields can back a call only through how it moved since the previous close, never through its level alone.
8 Oct 2026
Method v2
by hand, before the weekly audits
v2 started Each call also gets statistics from 15 years of hourly prices, such as how often the pair rose in this window after a similar move.
5 Oct 2026
Method v1
by hand, before the weekly audits
v1 started The first version: before each session the AI reads the morning news and calls each pair up or down, with no price statistics.

Locked in — window not closed yet

6 forecasts

These forecasts are already made, but you can't see them yet - only each one's hash: a 64-character code computed from the forecast and a random secret. Change one letter of the forecast and the hash comes out completely different, so we can't swap a forecast after the price moves. Copy a hash now and compare it after the result. How to check.

Tue 17 Nov · US · EUR/CHF a2ed6ee54f7670f3615b80a6dc2b184a7c53b02933834c69848da16f7a61144a .ots proof sealed 11:17 UTC · OpenTimestamps: submitted 17 Nov 11:17 UTC
Tue 17 Nov · US · EUR/USD dc71df323eb6e5ebb77c81841fe647c610475a643bfc912542358751f748efa1 .ots proof sealed 11:17 UTC · OpenTimestamps: submitted 17 Nov 11:17 UTC
Tue 17 Nov · US · USD/CHF c7e2e94762ef5591274dfcaa776ce5ac9616ae7c99823593d137aab821d21338 .ots proof sealed 11:17 UTC · OpenTimestamps: submitted 17 Nov 11:17 UTC
Tue 17 Nov · EU · EUR/CHF ebe94c226f20fadf8bf8f80c1f2b89900a430c27243eb2284ea3fdc27968c7bb .ots proof sealed 07:37 UTC · OpenTimestamps: submitted 17 Nov 07:37 UTC
Tue 17 Nov · EU · EUR/USD e8f568706266f88464e22942aae49acf7a73c1a8f471c8506c2d7ae4ab181d04 .ots proof sealed 07:37 UTC · OpenTimestamps: submitted 17 Nov 07:37 UTC
Tue 17 Nov · EU · USD/CHF 1b8b6fcf7d748f52c0cd7d96b6df310eb9218dd580c0a4a796d76b96779e4156 .ots proof sealed 07:37 UTC · OpenTimestamps: submitted 17 Nov 07:37 UTC

31 out of 171 shown

"Why it moved" is written by the AI after the window closed, from the day's news. It can be wrong, and it says so: clear, plausible or unclear is its own confidence in the explanation. Moves are in pips from the window's open.

Mon 5 Oct
Day session · v1
USD/CHF
▼ Down 51% sure
✗ Miss
+13.2 pips
max up +34.1 · max down -4.2
Why this forecast
The dollar had risen three weeks running and looked stretched, while worries in Europe sent money into the Swiss franc. Rising US interest rates limited how far it could fall.
Model's technical note Dollar overextended after three weekly gains and franc safe-haven bid on euro-area stress; rising US yields limit downside
Why it moved plausible
Franc gave back its French-stress safe-haven gains while 10y US yields near 5.3% kept the dollar firm
← NewerPage 2 of 2

Check a forecast yourself — 3 steps

A hash is a fingerprint of a text: the same text always gives the same 64 characters, and changing a single letter gives a completely different one. So if the hash you saw before the window matches the forecast shown after it, the forecast was not changed.

1 Before the window: copy the hash "Locked in" at the top of this page lists each upcoming forecast with only its hash. Press "Copy hash" and keep it (a note, a screenshot). 7b147f2ee27226a91593d4cf033e0524…
2 After the result: open the forecast Results arrive after the evening analysis (about 22:30 UTC). Find the forecast under "Every forecast" and click "Sealed before the window" under its result. It shows the forecast written as one line. fx2owl-fx-v1|2026-11-09|eu|EURUSD|up|55|secret
3 Check that they match Press "Check: hash the line in my browser" (nothing is sent to us), or paste the line into any SHA-256 tool. The result must be the hash you copied in step 1. ✓ Matches

Forecasts made before 9 Nov 2026 have no hash. The secret at the end of the line is random and shown only after the result - without it nobody could check the hash, and with it anyone can.

When was the hash made? Check it on opentimestamps.org

Each hash is also sent to OpenTimestamps, a free public service that writes it into the Bitcoin blockchain within a few hours. Nobody - us included - can change a Bitcoin block afterwards, so this proves the forecast was not rewritten after that block.

  1. Open the forecast's "Sealed before the window" and download two files: Download as .txt (the line) and Download the timestamp proof (.ots).
  2. Go to opentimestamps.org. In "Stamp & Verify", drop the .ots file.
  3. When it asks for the stamped file, drop the .txt file. It is hashed in your browser, not uploaded.
  4. It shows the Bitcoin block and its date - the hash existed by then. In the first few hours after sealing it shows the proof as pending: try again later.

The "sealed" time next to each forecast is when our server received the hash; the block's date is independent of us but comes later, usually one to a few hours after sealing.

For developers: the command line

The hashed line is fx2owl-fx-v1|date|session|PAIR|call|confidence|salt, UTF-8, no newline at the end. The salt is 32 random hex characters, kept secret until the window is scored, so the hash can't be guessed from the few possible forecasts. To hash a line yourself:

printf '%s' 'fx2owl-fx-v1|2026-11-06|eu|EURUSD|up|55|00112233445566778899aabbccddeeff' | sha256sum

(on macOS use shasum -a 256). On Windows PowerShell:

$t = 'fx2owl-fx-v1|2026-11-06|eu|EURUSD|up|55|00112233445566778899aabbccddeeff'
-join ([Security.Cryptography.SHA256]::Create().ComputeHash([Text.Encoding]::UTF8.GetBytes($t)) | ForEach-Object { $_.ToString('x2') })

The output must equal the published hash. With the OpenTimestamps client, put the downloaded .txt and .ots in one folder (their names already match) and run ots verify fx2owl-<hash>.txt.ots; ots upgrade on the same file saves the completed proof. The client verifies against your own Bitcoin Core node (a pruned one is fine).

Updated 17 Nov 2026 11:45 UTC. Source links go only to independent news sites and official sources (central banks, statistics offices). Educational content, not financial advice. Past results don't guarantee future ones.