Petro4Cast · Causal Forecasting

Best causal run per fuel, against the benchmark

Blue is actual demand, green is the baseline forecast (drivers held at their last value) with its 95% interval. Scenario percentages are each branch's largest deviation from the baseline inside its horizon.

12 weeks24 weeks

Gasoline0.783 MASE

Beats the benchmark (SARIMAX, MASE 0.848) by +7.6%

Next 12 weeks: average 8,733 kb/d↓ 2.7% vs last actual

Peak vs baseline →Oil Shock: -0.4%Refinery Down: +0.1%Slowdown: +0.9%

Gasoline0.686 MASE

Beats the benchmark (SARIMAX, MASE 0.717) by +4.3%

Next 24 weeks: average 8,602 kb/d↓ 4.3% vs last actual

Peak vs baseline →Oil Shock: -0.4%Refinery Down: +1.1%Slowdown: +0.9%

All runs, drivers and scenarios →

12 weeks24 weeks

Distillate (diesel)0.829 MASE

Beats the benchmark (SARIMAX, MASE 0.844) by +1.9%

Next 12 weeks: average 3,969 kb/d↑ 1.1% vs last actual

Peak vs baseline →Oil Shock: +0.2%Refinery Down: -3.4%Slowdown: -3.2%

Distillate (diesel)0.833 MASE

Beats the benchmark (SARIMAX, MASE 0.836) by +0.3%

Next 24 weeks: average 3,938 kb/d↑ 2.9% vs last actual

Peak vs baseline →Oil Shock: +0.2%Refinery Down: -2.7%Slowdown: -3.2%

All runs, drivers and scenarios →

12 weeks24 weeks

Jet fuel0.774 MASE

Beats the benchmark (ETS, MASE 0.792) by +2.3%

Next 12 weeks: average 1,708 kb/d↑ 3.3% vs last actual

Peak vs baseline →Oil Shock: +0.3%Refinery Down: -2.3%Slowdown: -2.7%

Jet fuel0.725 MASE

Beats the benchmark (SARIMAX, MASE 0.827) by +12.3%

Next 24 weeks: average 1,736 kb/d↑ 1.9% vs last actual

Peak vs baseline →Oil Shock: +0.3%Refinery Down: -4.0%Slowdown: -2.7%

All runs, drivers and scenarios →

All runs, drivers and scenarios per fuel

Same ten windows and horizons as the benchmark. MASE is scored with drivers held at their last value, so it is directly comparable with the benchmark; the known-drivers score is the upper bound.

GasolineDistillate (diesel)Jet fuel

12-week horizon24-week horizon

Causal SARIMAX, trained from 2005, 12-week horizon: test MASE 0.783 with drivers held (known: 0.720).
Beats the benchmark (SARIMAX, MASE 0.848) by +7.6%. Beats seasonal-naive by 22%.
p=1 d=1 q=1 K=6 trend=n log=True · drivers: p_gas_l8, vmt_l0, ev_demand_l0, ev_supply_l0

All runs, 12 weeks

Window	Horizon	MASE (held)	MASE (known)	MASE (blend)	MAPE (held)	Benchmark MASE	vs benchmark
2000	12w	0.785	0.776	0.778	3.07%	0.836 (SARIMAX)	+6.1%
2005	12w	0.783	0.720	0.780	3.10%	0.848 (SARIMAX)	+7.6%	best
2010	12w	0.787	0.745	0.775	3.11%	0.818 (SARIMAX)	+3.7%
2015	12w	0.854	0.842	0.839	3.33%	0.839 (SARIMAX)	-1.8%
2020	12w	0.850	0.856	0.863	3.37%	1.067 (ETS)	+20.3%

Best run: 2005, 12 weeks

ForecastDecompositionDriversScenarios

Drag to zoom, double-click to reset, click legend entries to hide series. Solid test line: future drivers held at their last value (honest). Dashed: actual future drivers fed in (upper bound).

Bars split each week's forecast change against the last actual week into seasonality, each driver family and the underlying level/trend; the line is the net change.

Driver coefficients

Driver	Fitted coefficient	What it means
X - Prices: retail gasoline price (8-week lag)	elasticity -0.028 (not significant)	A 10% rise in retail gasoline price (8-week lag) moves demand -0.3%.
X - Activity: vehicle miles travelled (mobility) (same week)	elasticity +0.449 (significant)	A 10% rise in vehicle miles travelled (mobility) (same week) moves demand +4.5%.
X - Events: demand-shock weeks (9/11, GFC, COVID) (same week)	-4.9% in flagged weeks (significant)	Demand in demand-shock weeks (9/11, GFC, COVID) runs -4.9% vs a normal week.
X - Events: supply-disruption weeks (hurricanes, freezes, outages) (same week)	-0.0% in flagged weeks (not significant)	Demand in supply-disruption weeks (hurricanes, freezes, outages) runs -0.0% vs a normal week.

What each model component contributes

Component	Fitted values	What it explains
AR - autoregressive	`p=1 (coef +0.31)`	Momentum: this week's forecast leans on the previous 1 week of demand with these weights.
I - integrated / trend	`d=1`	The model forecasts week-over-week changes, so the latest level is the baseline.
MA - moving average	`q=1 (coef -0.90)`	Shock smoothing: forecast errors from the last 1 week adjust this week's forecast.
S - seasonality (via X)	`K=6 Fourier pairs, 5.0% peak-to-trough`	Repeating yearly cycle - the calendar's contribution to demand.
X - Prices: retail gasoline price (8-week lag)	`elasticity -0.028 (not significant)`	A 10% rise in retail gasoline price (8-week lag) moves demand -0.3%.
X - Activity: vehicle miles travelled (mobility) (same week)	`elasticity +0.449 (significant)`	A 10% rise in vehicle miles travelled (mobility) (same week) moves demand +4.5%.
X - Events: demand-shock weeks (9/11, GFC, COVID) (same week)	`-4.9% in flagged weeks (significant)`	Demand in demand-shock weeks (9/11, GFC, COVID) runs -4.9% vs a normal week.
X - Events: supply-disruption weeks (hurricanes, freezes, outages) (same week)	`-0.0% in flagged weeks (not significant)`	Demand in supply-disruption weeks (hurricanes, freezes, outages) runs -0.0% vs a normal week.

Causal effects from the DAG (DoWhy)

Estimand	Adjust for (backdoor)	Coefficient	Effect of the scenario shock	Refutations (should be near 0 / unchanged)	Sensitivity
WTI +25% -> gasoline demand (8-week lag)	`crude_prod, crude_stk, macro`	-0.0162 (p=0.003)	-0.36%	placebo +0.000 · subset -0.017 · +cause -0.016	[-0.154, -0.018] · 5 valid sets: [-0.016, -0.016]
IPI -5% -> gasoline demand (4-week lag)	`ev_demand`	-0.1721 (p=0.668)	+0.89%	placebo +0.000 · subset -0.169 · +cause -0.173	[-1.133, -0.225] · 3 valid sets: [-0.172, -0.172]
utilization -10 pts -> gasoline output	`(none needed)`	+0.0034 (p=0.000)	-3.37%	placebo -0.000 · subset +0.003 · +cause +0.003	[+0.003, +0.004] · 10 valid sets: [+0.003, +0.003]

Coefficients are elasticities on year-on-year log changes (utilization in points), COVID window excluded. Placebo should be near 0; subset and random-cause should stay close to the estimate; the sweep shows the estimate's range under small unobserved confounders.

Same fitted model, four driver paths. Only the drivers change along each scenario's DAG chain.

Branch	What changes	Peak effect vs baseline	Week 12: forecast → effect
Oil Shock	WTI +25% -> gasoline demand (8-week lag): DoWhy effect -0.36% (adjusted for crude_prod, crude_stk, macro), phased in over 8 weeks	-32 (-0.36%) in week 11	8,421 → -30 (-0.36%)
Refinery Down	utilization -10 pts -> supply-event flag on for 2 weeks, crack -3.4 $/bbl, p_gas -2.9%, p_dsl -3.6%, p_jet -5.6%	+7 (+0.08%) in week 9	8,451 → +0 (+0.00%)
Slowdown	IPI -5% -> gasoline demand (4-week lag): DoWhy effect +0.89% (adjusted for ev_demand), phased in over 4 weeks	+78 (+0.89%) in week 11	8,526 → +75 (+0.89%)

Other runs

2000 · MASE 0.785 (drivers known 0.776)

p=3 d=1 q=2 K=6 trend=n log=False

2010 · MASE 0.787 (drivers known 0.745)

p=2 d=1 q=2 K=4 trend=n log=True

2015 · MASE 0.854 (drivers known 0.842)

p=3 d=1 q=2 K=6 trend=n log=False

2020 · MASE 0.850 (drivers known 0.856)

p=2 d=1 q=2 K=1 trend=t log=False

Causal SARIMAX, trained from 2000, 24-week horizon: test MASE 0.686 with drivers held (known: 0.701).
Beats the benchmark (SARIMAX, MASE 0.717) by +4.3%. Beats seasonal-naive by 31%.
p=3 d=1 q=2 K=6 trend=n log=False · drivers: p_gas_l8, vmt_l0, ev_demand_l0, ev_supply_l0

All runs, 24 weeks

Window	Horizon	MASE (held)	MASE (known)	MASE (blend)	MAPE (held)	Benchmark MASE	vs benchmark
2000	24w	0.686	0.701	0.692	2.79%	0.717 (SARIMAX)	+4.3%	best
2005	24w	0.759	0.710	0.728	3.08%	0.722 (SARIMAX)	-5.1%
2010	24w	0.743	0.698	0.722	3.01%	0.743 (SARIMAX)	-0.1%
2015	24w	0.784	0.705	0.741	3.17%	0.794 (SARIMAX)	+1.2%
2020	24w	0.765	0.749	0.818	3.10%	0.789 (SARIMAX)	+3.0%

Best run: 2000, 24 weeks

ForecastDecompositionDriversScenarios

Drag to zoom, double-click to reset, click legend entries to hide series. Solid test line: future drivers held at their last value (honest). Dashed: actual future drivers fed in (upper bound).

Bars split each week's forecast change against the last actual week into seasonality, each driver family and the underlying level/trend; the line is the net change.

Driver coefficients

Driver	Fitted coefficient	What it means
X - Prices: retail gasoline price (8-week lag)	+401.4 kbd per log-unit (significant)	A 10% rise in retail gasoline price (8-week lag) moves demand +38 kbd.
X - Activity: vehicle miles travelled (mobility) (same week)	+1539.7 kbd per log-unit (significant)	A 10% rise in vehicle miles travelled (mobility) (same week) moves demand +147 kbd.
X - Events: demand-shock weeks (9/11, GFC, COVID) (same week)	+8 kbd in flagged weeks (not significant)	Demand in demand-shock weeks (9/11, GFC, COVID) runs +8 kbd vs a normal week.
X - Events: supply-disruption weeks (hurricanes, freezes, outages) (same week)	+98 kbd in flagged weeks (not significant)	Demand in supply-disruption weeks (hurricanes, freezes, outages) runs +98 kbd vs a normal week.

What each model component contributes

Component	Fitted values	What it explains
AR - autoregressive	`p=3 (coef +0.98, -0.10, +0.01)`	Momentum: this week's forecast leans on the previous 3 weeks of demand with these weights.
I - integrated / trend	`d=1`	The model forecasts week-over-week changes, so the latest level is the baseline.
MA - moving average	`q=2 (coef -1.60, +0.61)`	Shock smoothing: forecast errors from the last 2 weeks adjust this week's forecast.
S - seasonality (via X)	`K=6 Fourier pairs, 563 kbd peak-to-trough`	Repeating yearly cycle - the calendar's contribution to demand.
X - Prices: retail gasoline price (8-week lag)	`+401.4 kbd per log-unit (significant)`	A 10% rise in retail gasoline price (8-week lag) moves demand +38 kbd.
X - Activity: vehicle miles travelled (mobility) (same week)	`+1539.7 kbd per log-unit (significant)`	A 10% rise in vehicle miles travelled (mobility) (same week) moves demand +147 kbd.
X - Events: demand-shock weeks (9/11, GFC, COVID) (same week)	`+8 kbd in flagged weeks (not significant)`	Demand in demand-shock weeks (9/11, GFC, COVID) runs +8 kbd vs a normal week.
X - Events: supply-disruption weeks (hurricanes, freezes, outages) (same week)	`+98 kbd in flagged weeks (not significant)`	Demand in supply-disruption weeks (hurricanes, freezes, outages) runs +98 kbd vs a normal week.

Causal effects from the DAG (DoWhy)

Estimand	Adjust for (backdoor)	Coefficient	Effect of the scenario shock	Refutations (should be near 0 / unchanged)	Sensitivity
WTI +25% -> gasoline demand (8-week lag)	`crude_prod, crude_stk, macro`	-0.0162 (p=0.003)	-0.36%	placebo +0.000 · subset -0.017 · +cause -0.016	[-0.154, -0.018] · 5 valid sets: [-0.016, -0.016]
IPI -5% -> gasoline demand (4-week lag)	`ev_demand`	-0.1721 (p=0.668)	+0.89%	placebo +0.000 · subset -0.169 · +cause -0.173	[-1.133, -0.225] · 3 valid sets: [-0.172, -0.172]
utilization -10 pts -> gasoline output	`(none needed)`	+0.0034 (p=0.000)	-3.37%	placebo -0.000 · subset +0.003 · +cause +0.003	[+0.003, +0.004] · 10 valid sets: [+0.003, +0.003]

Coefficients are elasticities on year-on-year log changes (utilization in points), COVID window excluded. Placebo should be near 0; subset and random-cause should stay close to the estimate; the sweep shows the estimate's range under small unobserved confounders.

Same fitted model, four driver paths. Only the drivers change along each scenario's DAG chain.

Branch	What changes	Peak effect vs baseline	Week 24: forecast → effect
Oil Shock	WTI +25% -> gasoline demand (8-week lag): DoWhy effect -0.36% (adjusted for crude_prod, crude_stk, macro), phased in over 8 weeks	-32 (-0.36%) in week 11	8,286 → -30 (-0.36%)
Refinery Down	utilization -10 pts -> supply-event flag on for 2 weeks, crack -3.4 $/bbl, p_gas -2.9%, p_dsl -3.6%, p_jet -5.6%	+98 (+1.10%) in week 1	8,316 → +0 (+0.00%)
Slowdown	IPI -5% -> gasoline demand (4-week lag): DoWhy effect +0.89% (adjusted for ev_demand), phased in over 4 weeks	+78 (+0.89%) in week 11	8,390 → +74 (+0.89%)

Other runs

2005 · MASE 0.759 (drivers known 0.710)

p=2 d=1 q=2 K=4 trend=c log=True

2010 · MASE 0.743 (drivers known 0.698)

p=2 d=1 q=2 K=4 trend=c log=True

2015 · MASE 0.784 (drivers known 0.705)

p=3 d=1 q=3 K=4 trend=c log=True

2020 · MASE 0.765 (drivers known 0.749)

p=3 d=1 q=3 K=2 trend=c log=True

12-week horizon24-week horizon

Causal SARIMAX, trained from 2020, 12-week horizon: test MASE 0.829 with drivers held (known: 0.841).
Beats the benchmark (SARIMAX, MASE 0.844) by +1.9%. Beats seasonal-naive by 17%.
p=2 d=1 q=3 K=3 trend=c log=True · drivers: hdd_anom_l0, ev_demand_l0, ev_supply_l0, pmi_l13

All runs, 12 weeks

Window	Horizon	MASE (held)	MASE (known)	MASE (blend)	MAPE (held)	Benchmark MASE	vs benchmark
2000	12w	0.872	0.884	0.871	6.97%	0.863 (SARIMAX)	-1.1%
2005	12w	0.931	0.938	0.928	7.44%	0.840 (SARIMAX)	-10.8%
2010	12w	0.933	0.947	0.922	7.45%	0.882 (SARIMAX)	-5.8%
2015	12w	0.897	0.908	0.884	7.16%	0.867 (SARIMAX)	-3.5%
2020	12w	0.829	0.841	0.843	6.59%	0.844 (SARIMAX)	+1.9%	best

Best run: 2020, 12 weeks

ForecastDecompositionDriversScenarios

Drag to zoom, double-click to reset, click legend entries to hide series. Solid test line: future drivers held at their last value (honest). Dashed: actual future drivers fed in (upper bound).

Bars split each week's forecast change against the last actual week into seasonality, each driver family and the underlying level/trend; the line is the net change.

Driver coefficients

Driver	Fitted coefficient	What it means
X - Weather: heating-degree-day anomaly (same week)	+0.0413% per unit (not significant)	Each unit of heating-degree-day anomaly (same week) moves demand +0.041%.
X - Events: demand-shock weeks (9/11, GFC, COVID) (same week)	-8.3% in flagged weeks (significant)	Demand in demand-shock weeks (9/11, GFC, COVID) runs -8.3% vs a normal week.
X - Events: supply-disruption weeks (hurricanes, freezes, outages) (same week)	-3.4% in flagged weeks (not significant)	Demand in supply-disruption weeks (hurricanes, freezes, outages) runs -3.4% vs a normal week.
X - Activity: manufacturing PMI (leading) (13-week lag)	elasticity +0.256 (significant)	A 10% rise in manufacturing PMI (leading) (13-week lag) moves demand +2.6%.

What each model component contributes

Component	Fitted values	What it explains
AR - autoregressive	`p=2 (coef -0.09, -0.73)`	Momentum: this week's forecast leans on the previous 2 weeks of demand with these weights.
I - integrated / trend	`d=1, intercept -4.45e-07`	The model forecasts week-over-week changes, so the latest level is the baseline. A fitted baseline trend is added on top.
MA - moving average	`q=3 (coef -0.81, +0.56, -0.74)`	Shock smoothing: forecast errors from the last 3 weeks adjust this week's forecast.
S - seasonality (via X)	`K=3 Fourier pairs, 11.1% peak-to-trough`	Repeating yearly cycle - the calendar's contribution to demand.
X - Weather: heating-degree-day anomaly (same week)	`+0.0413% per unit (not significant)`	Each unit of heating-degree-day anomaly (same week) moves demand +0.041%.
X - Events: demand-shock weeks (9/11, GFC, COVID) (same week)	`-8.3% in flagged weeks (significant)`	Demand in demand-shock weeks (9/11, GFC, COVID) runs -8.3% vs a normal week.
X - Events: supply-disruption weeks (hurricanes, freezes, outages) (same week)	`-3.4% in flagged weeks (not significant)`	Demand in supply-disruption weeks (hurricanes, freezes, outages) runs -3.4% vs a normal week.
X - Activity: manufacturing PMI (leading) (13-week lag)	`elasticity +0.256 (significant)`	A 10% rise in manufacturing PMI (leading) (13-week lag) moves demand +2.6%.

Causal effects from the DAG (DoWhy)

Estimand	Adjust for (backdoor)	Coefficient	Effect of the scenario shock	Refutations (should be near 0 / unchanged)	Sensitivity
WTI +25% -> distillate demand (8-week lag)	`crude_prod, crude_stk, macro`	+0.0106 (p=0.560)	+0.24%	placebo +0.002 · subset +0.012 · +cause +0.011	[-0.103, +0.006] · 5 valid sets: [+0.011, +0.011]
IPI -5% -> distillate demand (4-week lag)	`(none needed)`	+0.6295 (p=0.012)	-3.18%	placebo +0.011 · subset +0.632 · +cause +0.628	[-0.853, +0.220] · 24 valid sets: [+0.629, +0.629]
utilization -10 pts -> distillate output	`(none needed)`	+0.0098 (p=0.000)	-9.31%	placebo +0.000 · subset +0.010 · +cause +0.010	[+0.009, +0.010] · 10 valid sets: [+0.010, +0.010]

Coefficients are elasticities on year-on-year log changes (utilization in points), COVID window excluded. Placebo should be near 0; subset and random-cause should stay close to the estimate; the sweep shows the estimate's range under small unobserved confounders.

Same fitted model, four driver paths. Only the drivers change along each scenario's DAG chain.

Branch	What changes	Peak effect vs baseline	Week 12: forecast → effect
Oil Shock	WTI +25% -> distillate demand (8-week lag): DoWhy effect +0.24% (adjusted for crude_prod, crude_stk, macro), phased in over 8 weeks	+10 (+0.24%) in week 9	4,007 → +10 (+0.24%)
Refinery Down	utilization -10 pts -> supply-event flag on for 2 weeks, crack -3.4 $/bbl, p_gas -2.9%, p_dsl -3.6%, p_jet -5.6%	-132 (-3.42%) in week 2	3,998 → +0 (+0.00%)
Slowdown	IPI -5% -> distillate demand (4-week lag): DoWhy effect -3.18% (adjusted for (none needed)), phased in over 4 weeks	-130 (-3.18%) in week 9	3,871 → -127 (-3.18%)

Other runs

2000 · MASE 0.872 (drivers known 0.884)

p=2 d=1 q=3 K=3 trend=c log=True

2005 · MASE 0.931 (drivers known 0.938)

p=2 d=1 q=3 K=4 trend=n log=False

2010 · MASE 0.933 (drivers known 0.947)

p=2 d=1 q=3 K=3 trend=n log=False

2015 · MASE 0.897 (drivers known 0.908)

p=2 d=1 q=3 K=4 trend=n log=False

Causal SARIMAX, trained from 2020, 24-week horizon: test MASE 0.833 with drivers held (known: 0.843).
Beats the benchmark (SARIMAX, MASE 0.836) by +0.3%. Beats seasonal-naive by 17%.
p=0 d=1 q=1 K=3 trend=n log=True · drivers: hdd_anom_l0, ev_demand_l0, ev_supply_l0, pmi_l13

All runs, 24 weeks

Window	Horizon	MASE (held)	MASE (known)	MASE (blend)	MAPE (held)	Benchmark MASE	vs benchmark
2000	24w	0.847	0.855	0.849	6.85%	0.819 (SARIMAX)	-3.4%
2005	24w	0.883	0.890	0.883	7.17%	0.836 (SARIMAX)	-5.6%
2010	24w	0.896	0.907	0.893	7.27%	0.857 (SARIMAX)	-4.6%
2015	24w	0.871	0.881	0.863	7.06%	0.891 (ETS)	+2.3%
2020	24w	0.833	0.843	0.839	6.73%	0.836 (SARIMAX)	+0.3%	best

Best run: 2020, 24 weeks

ForecastDecompositionDriversScenarios

Drag to zoom, double-click to reset, click legend entries to hide series. Solid test line: future drivers held at their last value (honest). Dashed: actual future drivers fed in (upper bound).

Bars split each week's forecast change against the last actual week into seasonality, each driver family and the underlying level/trend; the line is the net change.

Driver coefficients

Driver	Fitted coefficient	What it means
X - Weather: heating-degree-day anomaly (same week)	+0.0451% per unit (not significant)	Each unit of heating-degree-day anomaly (same week) moves demand +0.045%.
X - Events: demand-shock weeks (9/11, GFC, COVID) (same week)	-8.3% in flagged weeks (significant)	Demand in demand-shock weeks (9/11, GFC, COVID) runs -8.3% vs a normal week.
X - Events: supply-disruption weeks (hurricanes, freezes, outages) (same week)	-2.7% in flagged weeks (not significant)	Demand in supply-disruption weeks (hurricanes, freezes, outages) runs -2.7% vs a normal week.
X - Activity: manufacturing PMI (leading) (13-week lag)	elasticity +0.260 (significant)	A 10% rise in manufacturing PMI (leading) (13-week lag) moves demand +2.6%.

What each model component contributes

Component	Fitted values	What it explains
AR - autoregressive	`p=0`	Not used: no direct carry-over from recent weeks.
I - integrated / trend	`d=1`	The model forecasts week-over-week changes, so the latest level is the baseline.
MA - moving average	`q=1 (coef -1.00)`	Shock smoothing: forecast errors from the last 1 week adjust this week's forecast.
S - seasonality (via X)	`K=3 Fourier pairs, 11.0% peak-to-trough`	Repeating yearly cycle - the calendar's contribution to demand.
X - Weather: heating-degree-day anomaly (same week)	`+0.0451% per unit (not significant)`	Each unit of heating-degree-day anomaly (same week) moves demand +0.045%.
X - Events: demand-shock weeks (9/11, GFC, COVID) (same week)	`-8.3% in flagged weeks (significant)`	Demand in demand-shock weeks (9/11, GFC, COVID) runs -8.3% vs a normal week.
X - Events: supply-disruption weeks (hurricanes, freezes, outages) (same week)	`-2.7% in flagged weeks (not significant)`	Demand in supply-disruption weeks (hurricanes, freezes, outages) runs -2.7% vs a normal week.
X - Activity: manufacturing PMI (leading) (13-week lag)	`elasticity +0.260 (significant)`	A 10% rise in manufacturing PMI (leading) (13-week lag) moves demand +2.6%.

Causal effects from the DAG (DoWhy)

Estimand	Adjust for (backdoor)	Coefficient	Effect of the scenario shock	Refutations (should be near 0 / unchanged)	Sensitivity
WTI +25% -> distillate demand (8-week lag)	`crude_prod, crude_stk, macro`	+0.0106 (p=0.560)	+0.24%	placebo +0.002 · subset +0.012 · +cause +0.011	[-0.103, +0.006] · 5 valid sets: [+0.011, +0.011]
IPI -5% -> distillate demand (4-week lag)	`(none needed)`	+0.6295 (p=0.012)	-3.18%	placebo +0.011 · subset +0.632 · +cause +0.628	[-0.853, +0.220] · 24 valid sets: [+0.629, +0.629]
utilization -10 pts -> distillate output	`(none needed)`	+0.0098 (p=0.000)	-9.31%	placebo +0.000 · subset +0.010 · +cause +0.010	[+0.009, +0.010] · 10 valid sets: [+0.010, +0.010]

Coefficients are elasticities on year-on-year log changes (utilization in points), COVID window excluded. Placebo should be near 0; subset and random-cause should stay close to the estimate; the sweep shows the estimate's range under small unobserved confounders.

Same fitted model, four driver paths. Only the drivers change along each scenario's DAG chain.

Branch	What changes	Peak effect vs baseline	Week 24: forecast → effect
Oil Shock	WTI +25% -> distillate demand (8-week lag): DoWhy effect +0.24% (adjusted for crude_prod, crude_stk, macro), phased in over 8 weeks	+10 (+0.24%) in week 8	4,076 → +10 (+0.24%)
Refinery Down	utilization -10 pts -> supply-event flag on for 2 weeks, crack -3.4 $/bbl, p_gas -2.9%, p_dsl -3.6%, p_jet -5.6%	-104 (-2.73%) in week 2	4,067 → +0 (+0.00%)
Slowdown	IPI -5% -> distillate demand (4-week lag): DoWhy effect -3.18% (adjusted for (none needed)), phased in over 4 weeks	-130 (-3.18%) in week 9	3,937 → -129 (-3.18%)

Other runs

2000 · MASE 0.847 (drivers known 0.855)

p=2 d=1 q=2 K=3 trend=c log=True

2005 · MASE 0.883 (drivers known 0.890)

p=2 d=1 q=2 K=3 trend=n log=False

2010 · MASE 0.896 (drivers known 0.907)

p=1 d=1 q=1 K=3 trend=c log=False

2015 · MASE 0.871 (drivers known 0.881)

p=2 d=1 q=3 K=3 trend=n log=False

12-week horizon24-week horizon

Causal SARIMAX, trained from 2010, 12-week horizon: test MASE 0.774 with drivers held (known: 0.747).
Beats the benchmark (ETS, MASE 0.792) by +2.3%. Beats seasonal-naive by 23%.
p=2 d=1 q=2 K=3 trend=n log=True · drivers: vmt_l0, emp_l4, ev_demand_l0, ev_supply_l0

All runs, 12 weeks

Window	Horizon	MASE (held)	MASE (known)	MASE (blend)	MAPE (held)	Benchmark MASE	vs benchmark
2000	12w	0.805	0.776	0.790	7.01%	0.783 (SARIMAX)	-2.8%
2005	12w	0.793	0.760	0.776	6.91%	0.773 (SARIMAX)	-2.6%
2010	12w	0.774	0.747	0.793	6.76%	0.792 (ETS)	+2.3%	best
2015	12w	0.800	0.753	0.822	7.00%	0.845 (ETS)	+5.4%
2020	12w	0.825	0.758	0.872	7.25%	0.825 (ETS)	-0.0%

Best run: 2010, 12 weeks

ForecastDecompositionDriversScenarios

Drag to zoom, double-click to reset, click legend entries to hide series. Solid test line: future drivers held at their last value (honest). Dashed: actual future drivers fed in (upper bound).

Bars split each week's forecast change against the last actual week into seasonality, each driver family and the underlying level/trend; the line is the net change.

Driver coefficients

Driver	Fitted coefficient	What it means
X - Activity: vehicle miles travelled (mobility) (same week)	elasticity +0.536 (significant)	A 10% rise in vehicle miles travelled (mobility) (same week) moves demand +5.4%.
X - Activity: employment level (macro) (4-week lag)	elasticity +3.864 (significant)	A 10% rise in employment level (macro) (4-week lag) moves demand +38.6%.
X - Events: demand-shock weeks (9/11, GFC, COVID) (same week)	-23.1% in flagged weeks (significant)	Demand in demand-shock weeks (9/11, GFC, COVID) runs -23.1% vs a normal week.
X - Events: supply-disruption weeks (hurricanes, freezes, outages) (same week)	-2.3% in flagged weeks (not significant)	Demand in supply-disruption weeks (hurricanes, freezes, outages) runs -2.3% vs a normal week.

What each model component contributes

Component	Fitted values	What it explains
AR - autoregressive	`p=2 (coef +0.15, -0.01)`	Momentum: this week's forecast leans on the previous 2 weeks of demand with these weights.
I - integrated / trend	`d=1`	The model forecasts week-over-week changes, so the latest level is the baseline.
MA - moving average	`q=2 (coef -1.19, +0.26)`	Shock smoothing: forecast errors from the last 2 weeks adjust this week's forecast.
S - seasonality (via X)	`K=3 Fourier pairs, 8.5% peak-to-trough`	Repeating yearly cycle - the calendar's contribution to demand.
X - Activity: vehicle miles travelled (mobility) (same week)	`elasticity +0.536 (significant)`	A 10% rise in vehicle miles travelled (mobility) (same week) moves demand +5.4%.
X - Activity: employment level (macro) (4-week lag)	`elasticity +3.864 (significant)`	A 10% rise in employment level (macro) (4-week lag) moves demand +38.6%.
X - Events: demand-shock weeks (9/11, GFC, COVID) (same week)	`-23.1% in flagged weeks (significant)`	Demand in demand-shock weeks (9/11, GFC, COVID) runs -23.1% vs a normal week.
X - Events: supply-disruption weeks (hurricanes, freezes, outages) (same week)	`-2.3% in flagged weeks (not significant)`	Demand in supply-disruption weeks (hurricanes, freezes, outages) runs -2.3% vs a normal week.

Causal effects from the DAG (DoWhy)

Estimand	Adjust for (backdoor)	Coefficient	Effect of the scenario shock	Refutations (should be near 0 / unchanged)	Sensitivity
WTI +25% -> jet demand (8-week lag)	`crude_prod, crude_stk, macro`	+0.0141 (p=0.109)	+0.32%	placebo -0.004 · subset +0.013 · +cause +0.014	[-0.114, +0.012] · 5 valid sets: [+0.014, +0.014]
IPI -5% -> jet demand (4-week lag)	`ev_demand`	+0.5371 (p=1.000)	-2.72%	placebo +0.028 · subset +0.551 · +cause +0.539	[-0.958, +0.350] · 3 valid sets: [+0.537, +0.537]
utilization -10 pts -> jet output	`(none needed)`	+0.0156 (p=0.000)	-14.47%	placebo +0.000 · subset +0.016 · +cause +0.016	[+0.015, +0.017] · 10 valid sets: [+0.016, +0.016]

Coefficients are elasticities on year-on-year log changes (utilization in points), COVID window excluded. Placebo should be near 0; subset and random-cause should stay close to the estimate; the sweep shows the estimate's range under small unobserved confounders.

Same fitted model, four driver paths. Only the drivers change along each scenario's DAG chain.

Branch	What changes	Peak effect vs baseline	Week 12: forecast → effect
Oil Shock	WTI +25% -> jet demand (8-week lag): DoWhy effect +0.32% (adjusted for crude_prod, crude_stk, macro), phased in over 8 weeks	+6 (+0.32%) in week 10	1,671 → +5 (+0.32%)
Refinery Down	utilization -10 pts -> supply-event flag on for 2 weeks, crack -3.4 $/bbl, p_gas -2.9%, p_dsl -3.6%, p_jet -5.6%	-41 (-2.29%) in week 1	1,666 → +0 (+0.00%)
Slowdown	IPI -5% -> jet demand (4-week lag): DoWhy effect -2.72% (adjusted for ev_demand), phased in over 4 weeks	-48 (-2.72%) in week 11	1,620 → -45 (-2.72%)

Other runs

2000 · MASE 0.805 (drivers known 0.776)

p=0 d=1 q=3 K=6 trend=n log=True

2005 · MASE 0.793 (drivers known 0.760)

p=0 d=1 q=1 K=5 trend=n log=True

2015 · MASE 0.800 (drivers known 0.753)

p=0 d=1 q=2 K=2 trend=n log=True

2020 · MASE 0.825 (drivers known 0.758)

p=0 d=1 q=2 K=2 trend=n log=True

Causal SARIMAX, trained from 2015, 24-week horizon: test MASE 0.725 with drivers held (known: 0.728).
Beats the benchmark (SARIMAX, MASE 0.827) by +12.3%. Beats seasonal-naive by 27%.
p=0 d=1 q=2 K=4 trend=n log=False · drivers: vmt_l0, emp_l4, ev_demand_l0, ev_supply_l0

All runs, 24 weeks

Window	Horizon	MASE (held)	MASE (known)	MASE (blend)	MAPE (held)	Benchmark MASE	vs benchmark
2000	24w	0.742	0.738	0.748	6.54%	0.779 (ETS)	+4.8%
2005	24w	0.737	0.739	0.744	6.49%	0.764 (ETS)	+3.6%
2010	24w	0.734	0.757	0.749	6.49%	0.768 (ETS)	+4.5%
2015	24w	0.725	0.728	0.763	6.39%	0.827 (SARIMAX)	+12.3%	best
2020	24w	0.769	0.738	0.832	6.80%	0.944 (SARIMAX)	+18.5%

Best run: 2015, 24 weeks

ForecastDecompositionDriversScenarios

Drag to zoom, double-click to reset, click legend entries to hide series. Solid test line: future drivers held at their last value (honest). Dashed: actual future drivers fed in (upper bound).

Bars split each week's forecast change against the last actual week into seasonality, each driver family and the underlying level/trend; the line is the net change.

Driver coefficients

Driver	Fitted coefficient	What it means
X - Activity: vehicle miles travelled (mobility) (same week)	+259.0 kbd per log-unit (significant)	A 10% rise in vehicle miles travelled (mobility) (same week) moves demand +25 kbd.
X - Activity: employment level (macro) (4-week lag)	+1585.1 kbd per log-unit (significant)	A 10% rise in employment level (macro) (4-week lag) moves demand +151 kbd.
X - Events: demand-shock weeks (9/11, GFC, COVID) (same week)	-357 kbd in flagged weeks (significant)	Demand in demand-shock weeks (9/11, GFC, COVID) runs -357 kbd vs a normal week.
X - Events: supply-disruption weeks (hurricanes, freezes, outages) (same week)	-73 kbd in flagged weeks (significant)	Demand in supply-disruption weeks (hurricanes, freezes, outages) runs -73 kbd vs a normal week.

What each model component contributes

Component	Fitted values	What it explains
AR - autoregressive	`p=0`	Not used: no direct carry-over from recent weeks.
I - integrated / trend	`d=1`	The model forecasts week-over-week changes, so the latest level is the baseline.
MA - moving average	`q=2 (coef -1.07, +0.22)`	Shock smoothing: forecast errors from the last 2 weeks adjust this week's forecast.
S - seasonality (via X)	`K=4 Fourier pairs, 141 kbd peak-to-trough`	Repeating yearly cycle - the calendar's contribution to demand.
X - Activity: vehicle miles travelled (mobility) (same week)	`+259.0 kbd per log-unit (significant)`	A 10% rise in vehicle miles travelled (mobility) (same week) moves demand +25 kbd.
X - Activity: employment level (macro) (4-week lag)	`+1585.1 kbd per log-unit (significant)`	A 10% rise in employment level (macro) (4-week lag) moves demand +151 kbd.
X - Events: demand-shock weeks (9/11, GFC, COVID) (same week)	`-357 kbd in flagged weeks (significant)`	Demand in demand-shock weeks (9/11, GFC, COVID) runs -357 kbd vs a normal week.
X - Events: supply-disruption weeks (hurricanes, freezes, outages) (same week)	`-73 kbd in flagged weeks (significant)`	Demand in supply-disruption weeks (hurricanes, freezes, outages) runs -73 kbd vs a normal week.

Causal effects from the DAG (DoWhy)

Estimand	Adjust for (backdoor)	Coefficient	Effect of the scenario shock	Refutations (should be near 0 / unchanged)	Sensitivity
WTI +25% -> jet demand (8-week lag)	`crude_prod, crude_stk, macro`	+0.0141 (p=0.109)	+0.32%	placebo -0.004 · subset +0.013 · +cause +0.014	[-0.114, +0.012] · 5 valid sets: [+0.014, +0.014]
IPI -5% -> jet demand (4-week lag)	`ev_demand`	+0.5371 (p=1.000)	-2.72%	placebo +0.028 · subset +0.551 · +cause +0.539	[-0.958, +0.350] · 3 valid sets: [+0.537, +0.537]
utilization -10 pts -> jet output	`(none needed)`	+0.0156 (p=0.000)	-14.47%	placebo +0.000 · subset +0.016 · +cause +0.016	[+0.015, +0.017] · 10 valid sets: [+0.016, +0.016]

Coefficients are elasticities on year-on-year log changes (utilization in points), COVID window excluded. Placebo should be near 0; subset and random-cause should stay close to the estimate; the sweep shows the estimate's range under small unobserved confounders.

Same fitted model, four driver paths. Only the drivers change along each scenario's DAG chain.

Branch	What changes	Peak effect vs baseline	Week 24: forecast → effect
Oil Shock	WTI +25% -> jet demand (8-week lag): DoWhy effect +0.32% (adjusted for crude_prod, crude_stk, macro), phased in over 8 weeks	+6 (+0.32%) in week 14	1,650 → +5 (+0.32%)
Refinery Down	utilization -10 pts -> supply-event flag on for 2 weeks, crack -3.4 $/bbl, p_gas -2.9%, p_dsl -3.6%, p_jet -5.6%	-73 (-3.98%) in week 1	1,644 → +0 (+0.00%)
Slowdown	IPI -5% -> jet demand (4-week lag): DoWhy effect -2.72% (adjusted for ev_demand), phased in over 4 weeks	-48 (-2.72%) in week 16	1,600 → -45 (-2.72%)

Other runs

2000 · MASE 0.742 (drivers known 0.738)

p=3 d=1 q=2 K=4 trend=n log=False

2005 · MASE 0.737 (drivers known 0.739)

p=3 d=1 q=2 K=5 trend=c log=False

2010 · MASE 0.734 (drivers known 0.757)

p=0 d=1 q=3 K=6 trend=n log=False

2020 · MASE 0.769 (drivers known 0.738)

p=1 d=1 q=3 K=1 trend=c log=False

Edge equations used by the scenarios

Slopes on year-on-year changes along each DAG edge, COVID window excluded.

Edge	Slope	Units	Status
`wti->p_gas`	0.5499	elasticity (log/log)	ok
`wti->p_dsl`	0.6109	elasticity (log/log)	ok
`wti->p_jet`	1.022	elasticity (log/log)	ok
`ipi->emp`	0.2436	elasticity (log/log)	ok
`ipi->vmt`	0.1035	elasticity (log/log)	ok
`util->crack`	0.3396	$/bbl per utilization point	ok
`crack->p_gas`	0.0088	log-price per $/bbl	ok
`crack->p_dsl`	0.0108	log-price per $/bbl	ok
`crack->p_jet`	0.0169	log-price per $/bbl	ok

Stage 03Causal forecast, drivers and benchmarking

What goes into this stage, what happens, and what comes out.

What goes in

The weekly panel and the DAG shortlist of drivers
Baseline forecasts for the same windows

What happens

Tune orders and driver lags with Optuna, with differencing constrained to at least one
Project future driver values with seasonal naive plus drift, since real forecasts do not know them
Blend the causal and univariate forecasts and score both against the benchmark
Decompose each forecast into seasonality, trend and driver contributions

What comes out

Weekly forecasts with intervals per fuel
A driver decomposition that says why the number moved
A benchmark table: score and windows won per fuel

SARIMAX + driversOptunaSeasonal naive + driftDecompositionBenchmark tableSee all four stages on the home page →