Diesel: cap vs reconstructed cost
- MINT published cap
- Reconstructed cost
Both series in €/litre on a single shared axis. The vertical gap is the unexplained margin.
Correlation by lag
Pearson r on day-to-day changes, testing the cap against the benchmark from N days earlier.
Read this chart with caution. A 4-day trailing average fits better than any single-day lag, so this peak most likely reflects the middle of an averaging window rather than a real reaction delay.
Diesel: projection
- Published
- Locked in
- Simulated (80% band)
The cap follows the market with a lag, so the sessions driving the next few days have already happened.
Market inputs known through 2026-09-22 — every benchmark session feeding those days is already observed, so they are arithmetic rather than a speculative forecast. After that the inputs themselves are unknown. After that the market itself is simulated as a random walk (2,000 paths, daily σ €0.0307/L), which is why the band flares. Even with a flat market the cap keeps moving through 2026-09-26 as old sessions roll out of the 4-day window.
Chance the cap moves
- Rise
- Unchanged
- Fall
Based on benchmark sessions that have already closed, not on a forecast of the oil market.
Accuracy is measured walk-forward on the one-day-ahead call: each day was predicted using only data available before it. On 49 test days that figure itself carries several points of error, so treat a small edge over the base rate as unproven.
Days without the locked in tag depend on market sessions that have not happened. Their probabilities come from simulating the benchmark forward, so they drift toward the base rate — that convergence is the model saying it does not know, not a prediction of stability.
Unexplained margin over time
Published cap minus reconstructed cost, €/litre. Watch the drift, not the level.
The level absorbs the benchmark basis and the transport estimate as well as any genuine unclaimed margin (mean €0.000/L), so it is not interpretable on its own. A sustained change in shape is the real signal.
Model statistics
- Benchmark
- NYMEX ULSD (HO=F)
- Best-fitting variant
- 4-day mean, 2-day lag
- Pass-through β
- 1.099 (vs 1.18 for full)
- r² on changes
- 0.620
- r on levels
- 0.982 — not trusted
- Margin σ (levels)
- €0.0373/L
- Forecast σ (changes)
- €0.0147/L
- Residual basis
- +€0.0003/L
What this cannot tell you
- 225 model variants were compared on 65 observations. The winner is flattered by that search — treat it as a hypothesis.
- NYMEX ULSD (HO=F) is a free proxy. Whatever reference MINT actually uses differs in delivery location and contract form, and that gap moves. It puts a ceiling on achievable r that no tuning will lift.
- Transport is an unverified estimate (€0.030/L). It shifts the margin level but, being constant, cannot affect the correlation or the lag.
- The projection scored 78% on 49 backtested days, 16 points above always guessing the commonest outcome. That gap is larger than the ~6-point sampling error, but the baseline itself is weak here — a longer run is still needed to trust it.
- 3 days missing from the MINT series. Changes spanning a gap of more than four days are excluded.