NY Harbor ULSD Crack Spread Futures

The tests show no consistent, material advantage over the no-change baseline.

Model: Agentic Investor TSFM (foundation model)

Accuracy vs naive

+6%Matches the no-change baseline over 10 tests (10 tests, 8 won)

Direction

70%No clear edge over a coin flip (84 of 120 calls right in 10 tests)

Turning points

12 of 56Anticipated some turns (56 turns in 10 tests, 0 without a turn)

Error vs baseline

6% lowerError ratio 0.94; 1.00 is a tie.

Tests won

8 / 10Complete rolling backtests; ties count as no win.

Test period

2016-11-01 to 2026-10-0112 observations ahead per test.

80% band coverage

78%Observed outcomes inside the band across all stored test windows.

• Band coverage uses all stored test windows, including model-selection windows when present; it does not determine the confidence label.
• Evaluated 2026-11-14 against the cached history through 2026-10-01.
• This is an evaluation of the serving model; it does not establish a winner across four model classes.