Structural Relationships Between Observable Fuel Oil Inventory Pressure and Forward Singapore HSFO Pricing Behaviour.
This paper presents the Maven Fuel Oil Inventory Pressure Research Framework, a systematic research programme designed to test whether observable physical fuel oil inventory pressure exhibits statistically meaningful relationships with future Singapore HSFO crack pricing behaviour.
The framework was developed to evaluate whether cargo-level physical intelligence, regional arrivals data, operational cargo maturity, and inventory accumulation behaviour can be reconstructed into a repeatable research architecture capable of explaining forward pricing structure.
The research does not attempt to present a deterministic trading model. Instead, it functions as a structural market analysis framework. Its purpose is to test whether observable supply pressure contains measurable information about future pricing behaviour, and to identify which definitions of physical inventory pressure are most informative.
The research focused on Singapore HSFO cracks because Singapore remains one of the most important regional pricing and clearing centres for fuel oil and bunker-related supply. The analysis therefore tested whether observable regional fuel oil arrivals, particularly HSFO-related cargoes, could help explain future movements in Singapore HSFO crack structure.
Across the research universe, the strongest observed Spearman relationship reached 0.6155, while the strongest observed Pearson relationship reached 0.5276. The strongest relationships consistently exhibited an inverse relationship between observable inventory pressure and future Singapore HSFO crack levels.
The concentration of strong results around similar parameter combinations suggests that the relationships identified were not isolated statistical outliers but instead reflected a coherent and repeatable market structure.
In practical terms, the results indicate that higher observable inventory pressure was associated with weaker future Singapore HSFO crack structure, while lower observable inventory conditions were associated with stronger future crack behaviour.
The research programme evaluated 22,572 valid experimental configurations across more than 4.31 million underlying observation rows.
The strongest results were generated by inventory accumulation measures rather than single-week inventory snapshots. Rolling and decayed inventory methodologies consistently outperformed prompt-only inventory signals, with the highest correlations emerging from four-to-six-week inventory accumulation windows.
The top-performing models were highly consistent. The strongest configurations repeatedly combined:
| Component | Strongest Pattern Observed |
|---|---|
| Product | HSFO-only |
| Confidence Filter | confidence_clean |
| ETA Window | 0 to 4 weeks |
| Stem Status | active_stem_only and confirmed_only |
| Geography | core_singapore_market and extended_bunker_supply_region |
| Signal Type | rolling_6w_inventory_sum and decayed_equal_6w_inventory |
| Target Type | crack_level |
| Lag | 2 weeks forward |
| Direction | Inverse relationship between inventory pressure and crack strength |
The best observed configuration produced:
| Metric | Result |
|---|---|
| Spearman | -0.6155 |
| Pearson | -0.5212 |
| Observation Count | 202 |
| Product | hsfo_only |
| Geography | extended_bunker_supply_region |
| Confidence | confidence_clean |
| Stem Status | active_stem_only |
| Signal Type | decayed_equal_6w_inventory |
| Target | target_crack_lag_2 |
| Lag | 2 weeks |
| ETA Window | 0 to 4 |
The best results were economically coherent. They suggest that observable HSFO inventory pressure behaves less like a short-term event signal and more like a broader market-state variable. Persistent inventory accumulation appears to carry substantially more information about future Singapore HSFO crack structure than a single weekly arrivals snapshot.
The top 100 strongest configurations were especially important because they showed concentrated and repeatable structure:
| Factor Level | Share of Top 100 Models |
|---|---|
| confidence_clean | 100% |
| hsfo_only | 100% |
| ETA window 0 to 4 | 100% |
| crack_level target | 100% |
| lag 2 | 74% |
| lag 3 | 26% |
| active_stem_only | 43% |
| confirmed_only | 43% |
| confirmed_plus_projected | 14% |
This concentration was observed simultaneously across multiple independent dimensions, including product selection, confidence filtering, ETA window construction, target type, lag structure, and stem status. The clustering of results around the same physical interpretation of the market suggests that the strongest findings are unlikely to be isolated statistical outliers and instead reflect a persistent underlying relationship between observable inventory pressure and forward HSFO pricing structure.
The objective of the framework was to determine whether observable physical fuel oil inventory pressure has a measurable relationship with future Singapore HSFO pricing behaviour.
The framework was designed to answer five core questions:
| Research Question | Purpose |
|---|---|
| Does observable inventory pressure matter? | Tests whether physical arrivals contain pricing information. |
| Which product definition works best? | Tests whether HSFO-only flows outperform broader HSFO-related groupings. |
| Which cargo certainty level matters most? | Tests whether confirmed or high-confidence cargoes produce cleaner signals. |
| Which timing window matters most? | Tests whether prompt and forward arrivals outperform broader historical windows. |
| Which signal construction method works best? | Tests whether rolling, decayed, binary, change, or level signals perform best. |
The broader institutional objective was to build a repeatable research system rather than a one-off correlation exercise.
This matters because a single strong correlation can be misleading. A robust research framework should show whether similar relationships persist across different configurations, filters, regions, products, signal types, and lag horizons.
The core economic hypothesis is that observable physical inventory pressure influences Singapore HSFO crack structure through regional supply balance.
The simplified logic chain is:
The expected relationship was inverse.
Higher observable inventory pressure should generally be associated with weaker future HSFO crack structure, because increased supply availability places pressure on prompt and nearby pricing.
Lower observable inventory pressure should generally be associated with stronger future HSFO crack structure, because tighter physical availability supports crack pricing.
The framework also tested whether this relationship appears immediately or with a lag. The strongest results emerging around the two-week forward horizon suggest that physical pressure may require time to propagate through regional balances, trading behaviour, bunker supply chains, blending systems, and benchmark pricing.
Fuel oil markets are shaped by physical balance dynamics that are not always fully visible through price data alone.
Key physical drivers include:
| Driver | Relevance |
|---|---|
| Cargo arrivals | Directly affect regional supply availability. |
| Stem status | Determines whether a cargo is projected, active, confirmed, or arrived. |
| Product classification | Determines whether cargoes are directly relevant to HSFO pricing. |
| Destination region | Determines whether cargoes are connected to the Singapore pricing ecosystem. |
| Arrival timing | Determines whether supply is prompt, deferred, or already confirmed. |
| Inventory persistence | Determines whether supply pressure is isolated or sustained. |
Traditional price analysis starts from the price itself. This research starts from observable physical market behaviour and tests whether it can explain future price structure.
That distinction is important. If physical flow intelligence contains useful information before it is fully reflected in benchmark pricing, then cargo-level observability can become a meaningful market-state indicator.
The research used cargo-level fuel oil intelligence aggregated into weekly inventory-pressure systems.
The framework tested:
| Metric | Result |
|---|---|
| Valid experimental results | 22,572 |
| Total observation rows | 4,313,381 |
| Average observations per valid result | 191.1 |
| Best observed absolute Spearman | 0.6155 |
| Best observed absolute Pearson | 0.5276 |
| 90th percentile absolute Spearman | 0.2060 |
The strongest Spearman and Pearson results were not necessarily generated by the same experimental configuration. The values above represent the highest absolute correlations observed across the full experimental universe.
The analysis used valid experiment results only. Results were filtered to remove insufficient observations, constant inputs, and invalid statistical outputs.
The experimental universe included multiple factor families:
| Factor | Levels Tested |
|---|---|
| Confidence | 3 |
| Geography | 3 |
| Product | 4 |
| Stem Status | 5 |
| ETA Window | 2 |
| Weighting Set | 5 |
| Signal Type | 21 |
| Target Type | 4 |
| Target Feature | 16 |
| Lag Weeks | 4 |
| Rolling Window | 5 |
| Decay Type | 4 |
A separate Fuel Oil Inventory Pressure Research Factor Catalogue defines every factor level used in the research universe and can be provided on request.
The framework used three primary data inputs.
The physical inventory side of the analysis was built from Maven Knowledge cargo-level fuel oil intelligence. This included observable cargo movements, product classifications, stem status, confidence classifications, destination geography, ETA timing, and reported cargo volumes.
The pricing side of the analysis used Singapore HSFO bunker pricing and Brent crude pricing. These two series were combined to construct a Singapore HSFO crack series, allowing the framework to evaluate the relationship between observable physical inventory pressure and relative HSFO pricing strength rather than outright fuel oil prices alone.
Using the crack rather than the outright HSFO price was important because it helps isolate fuel oil-specific supply and demand dynamics from broader movements in the crude oil market.
The research period covered 2022 through 2025. This provided approximately four years of weekly market observations across multiple market regimes, including periods of tighter and weaker fuel oil market structure. A small number of gaps within the weekly data series were filled to preserve continuity and maintain a consistent observation framework throughout the study period.
The analysis was deliberately constructed around a weekly cadence. Maven Knowledge publishes physical fuel oil intelligence on a weekly reporting cycle, so all inventory observations were aligned to Monday report dates. Singapore HSFO bunker prices, Brent prices, and the derived crack series were aligned to the same weekly Monday timestamps.
This produced approximately 200 weekly Maven inventory observations across the research period.
The Monday alignment was important for three reasons.
In practical terms, each weekly Maven report can be used to update the observable inventory pressure signal, compare it against the historical framework, and assess whether prevailing physical supply conditions are consistent with stronger or weaker forward Singapore HSFO crack structure.
The methodology consisted of five steps.
Cargo-level physical observations were grouped into weekly inventory measures.
Each inventory signal was defined by a combination of:
| Dimension | Example |
|---|---|
| Product | hsfo_only |
| Geography | core_singapore_market |
| Stem Status | confirmed_only |
| Confidence Filter | confidence_clean |
| ETA Window | 0 to 4 |
| Signal Type | rolling_6w_inventory_sum |
This created multiple plausible definitions of observable fuel oil inventory pressure.
Inventory signals were aligned against forward Singapore HSFO crack targets.
The framework tested crack levels, crack changes, z-scored crack levels, and z-scored crack changes across one-to-four-week forward lags.
The framework calculated both Pearson and Spearman correlations.
| Metric | Purpose |
|---|---|
| Pearson | Measures linear relationship strength. |
| Spearman | Measures monotonic rank relationship strength. |
Spearman was especially important because physical commodity markets often behave non-linearly. A market may respond differently once inventory pressure becomes extreme, persistent, or regime-defining.
Configurations were initially ranked using relationship strength. Observation counts and consistency across related experiments were then used as secondary validation measures when interpreting the results.
The strongest observed configurations were reviewed alongside broader factor-level performance to avoid over-interpreting isolated outliers.
The framework then assessed which factors and factor levels appeared most frequently in the strongest models.
This matters because repeatability is more important than a single winning result.
The strongest evidence came from factor levels that combined:
| Evidence Type | Why It Matters |
|---|---|
| High best Spearman | Shows maximum observed explanatory relationship. |
| High 90th percentile Spearman | Shows strength across better-performing configurations. |
| Frequent top-100 appearance | Shows consistency among strongest models. |
| Economic plausibility | Confirms the result makes market sense. |
The top 10 results were tightly clustered and economically consistent.
| Rank | Spearman | Pearson | Geography | Product | Status | Confidence | Signal | Target | Lag |
|---|---|---|---|---|---|---|---|---|---|
| 1 | -0.6155 | -0.5212 | extended_bunker_supply_region | hsfo_only | active_stem_only | confidence_clean | decayed_equal_6w_inventory | crack_level | 2 |
| 2 | -0.6155 | -0.5212 | extended_bunker_supply_region | hsfo_only | confirmed_only | confidence_clean | decayed_equal_6w_inventory | crack_level | 2 |
| 3 | -0.6154 | -0.5212 | extended_bunker_supply_region | hsfo_only | active_stem_only | confidence_clean | rolling_6w_inventory_sum | crack_level | 2 |
| 4 | -0.6154 | -0.5212 | extended_bunker_supply_region | hsfo_only | confirmed_only | confidence_clean | rolling_6w_inventory_sum | crack_level | 2 |
| 5 | -0.6152 | -0.5203 | core_singapore_market | hsfo_only | active_stem_only | confidence_clean | decayed_equal_6w_inventory | crack_level | 2 |
| 6 | -0.6152 | -0.5203 | core_singapore_market | hsfo_only | confirmed_only | confidence_clean | decayed_equal_6w_inventory | crack_level | 2 |
| 7 | -0.6151 | -0.5203 | core_singapore_market | hsfo_only | active_stem_only | confidence_clean | rolling_6w_inventory_sum | crack_level | 2 |
| 8 | -0.6151 | -0.5203 | core_singapore_market | hsfo_only | confirmed_only | confidence_clean | rolling_6w_inventory_sum | crack_level | 2 |
| 9 | -0.6107 | -0.5184 | broad_south_east_asia | hsfo_only | active_stem_only | confidence_clean | rolling_6w_inventory_sum | crack_level | 2 |
| 10 | -0.6107 | -0.5184 | broad_south_east_asia | hsfo_only | confirmed_only | confidence_clean | rolling_6w_inventory_sum | crack_level | 2 |
The top models show a clear and consistent pattern.
Many of the highest-ranked configurations differ only marginally in geography, stem status, or inventory construction methodology, yet produce almost identical correlation values. This suggests the relationship is robust to reasonable variations in model specification rather than being dependent on a single narrowly-defined model structure.
This consistency is one of the most important findings in the research.
The framework assessed how much each factor changed performance across its internal factor levels.
This is not presented as an average correlation. Instead, it is used as a diagnostic measure of factor separation: how much difference there was between the strongest and weakest levels within the same factor.
The chart below summarises factor separation across the full experimental universe. Larger separation indicates that the research results were more sensitive to that particular framework choice.
| Rank | Factor | Separation Score | Interpretation |
|---|---|---|---|
| 1 | Signal Type | 0.1554 | Signal construction was the most important design choice. |
| 2 | Stem Status | 0.0346 | Cargo operational maturity materially affected signal quality. |
| 3 | Confidence | 0.0340 | Higher-confidence filtering materially improved results. |
| 4 | Product | 0.0286 | Product definition mattered, with HSFO-only dominant in top models. |
| 5 | Rolling Window | 0.0265 | Rolling accumulation window influenced signal strength. |
| 6 | Decay Type | 0.0183 | Equal decay outperformed more complex decay structures. |
| 7 | ETA Window | 0.0132 | Forward prompt windows outperformed wider backward-looking windows. |
| 8 | Weighting Set | 0.0087 | Named weighting sets mattered less than rolling and decayed accumulation systems. |
| 9 | Geography | 0.0005 | Geography mattered less on average, although the top models still clustered around Singapore-linked regions. |
| Rank | Factor | Separation Score | Interpretation |
|---|---|---|---|
| 1 | Target Feature | 0.1493 | Crack level lag targets clearly outperformed crack change targets. |
| 2 | Target Type | 0.1264 | Crack levels were materially stronger than crack changes or z-scored changes. |
| 3 | Lag Weeks | 0.0157 | Two and three-week forward horizons dominated the strongest models. |
This ranking is important because it shows that the research result was not driven only by geography or product selection. The biggest design decision was how inventory pressure was constructed.
Rolling and decayed inventory accumulation systems consistently outperformed simple inventory definitions, reinforcing the view that persistent inventory pressure contains more explanatory information than single-period inventory observations.
Cargo-level physical intelligence contains varying degrees of certainty.
Some observations represent confirmed physical movement. Others include uncertainty around destination, product, routing, or voyage outcome.
The research therefore tested whether stricter cargo confidence filtering improves signal quality.
| Rank | Confidence Level | Best Spearman | P90 Spearman | Top 100 Share |
|---|---|---|---|---|
| 1 | confidence_clean | 0.6155 | 0.2449 | 100% |
| 2 | navigation_clean | 0.4773 | 0.1844 | 0% |
| 3 | base_clean | 0.4773 | 0.1853 | 0% |
The confidence_clean filter dominated the strongest results.
It produced the highest observed Spearman relationship, the strongest P90 Spearman result, and appeared in 100% of the top 100 configurations.
This strongly supports the hypothesis that cargo confidence and data quality matter. Removing uncertain or lower-quality cargo observations materially improved the structural relationship between observable inventory pressure and future Singapore HSFO crack behaviour.
For research and monitoring purposes, the highest-confidence cargo universe should be treated as the primary analytical dataset.
Broader datasets may remain useful for market context, but the strongest pricing relationships emerged from cleaner, higher-certainty physical flow data.
Singapore HSFO pricing may respond to different geographic definitions of supply pressure.
A narrow Singapore-focused region may capture direct local inventory conditions, while a broader South East Asian region may capture regional bunker supply, blending, and redistribution effects.
| Rank | Geography | Best Spearman | P90 Spearman | Top 100 Share |
|---|---|---|---|---|
| 1 | core_singapore_market | 0.6152 | 0.2081 | 35% |
| 2 | extended_bunker_supply_region | 0.6155 | 0.2055 | 37% |
| 3 | broad_south_east_asia | 0.6107 | 0.2045 | 28% |
Geography showed less separation than other factors. All three geography definitions produced strong results and all appeared in the top 100 models.
The results should not be interpreted as a single geography dominating the framework. The core_singapore_market ranked highest on the factor-level ranking and produced the strongest P90 Spearman result. The extended_bunker_supply_region produced the single best observed Spearman result and appeared most frequently in the top 100 configurations. The broad_south_east_asia definition also performed strongly, with only a modest drop in best Spearman and P90 Spearman.
This suggests that the strongest inventory relationship is not confined to one narrow geographic definition. Instead, Singapore HSFO pricing appears connected to a broader regional supply ecosystem.
The evidence supports treating Singapore as part of a wider regional fuel oil and bunker supply system rather than as an isolated local inventory point.
For ongoing monitoring, both core Singapore-linked arrivals and broader regional bunker supply should be tracked.
The framework tested whether Singapore HSFO pricing is best explained by pure HSFO cargo flows or by broader HSFO-related material, including SRFO and heavy sweet-related cargoes.
| Product Level | Best Spearman | P90 Spearman | Top 100 Share |
|---|---|---|---|
| hsfo_only | 0.6155 | 0.2103 | 100% |
| hsfo_hsweet | 0.3988 | 0.1387 | 0% |
| hsfo_srfo | 0.3820 | 0.2077 | 0% |
| hsfo_srfo_hsweet | 0.3768 | 0.2141 | 0% |
The product comparison produced one of the clearest findings in the entire research programme.
HSFO-only cargo definitions materially outperformed all broader HSFO-related product groupings, suggesting that pricing relationships were strongest when the inventory signal remained closely aligned with the underlying benchmark market.
HSFO-only appeared in 100% of the top 100 configurations and produced the strongest observed Spearman relationship by a substantial margin.
None of the broader product groupings appeared in the top 100 configurations. While some alternative groupings produced respectable factor-level results, they did not generate the strongest observed relationships.
This is an important finding because it suggests that direct HSFO physical flow observability was more informative than broader HSFO-adjacent product aggregation.
For Singapore HSFO crack analysis, HSFO-only physical inventory pressure should remain the primary signal definition.
Broader product groupings may still provide market context, but they should not replace HSFO-only analysis in the core framework.
Cargo status captures the operational maturity of a physical movement.
Projected cargoes may contain forward-looking information, but they also introduce uncertainty. Active and confirmed cargoes are more physically reliable and may therefore produce cleaner pricing relationships.
| Stem Status | Best Spearman | P90 Spearman | Top 100 Share |
|---|---|---|---|
| active_stem_only | 0.6155 | 0.2332 | 43% |
| confirmed_only | 0.6155 | 0.2332 | 43% |
| confirmed_plus_projected | 0.5912 | 0.2215 | 14% |
| projected_only | 0.5405 | 0.1493 | 0% |
A fifth stem-status category, anchored_only, was included within the broader experimental universe but did not generate sufficient valid statistical outputs to form part of the factor-level comparison presented here.
Active and confirmed cargoes clearly dominated the strongest results.
Together, active_stem_only and confirmed_only accounted for 86% of the top 100 configurations while producing the highest observed Spearman relationships and strongest P90 performance.
confirmed_plus_projected also appeared in the top 100, but at a much lower frequency. projected_only did not appear in the top 100 configurations.
This suggests that physical certainty matters. The market relationship was strongest when the inventory signal was built from cargoes with greater operational maturity and higher discharge confidence.
Confirmed and active voyage cargoes should form the core of the inventory pressure framework.
Projected cargoes may be useful for forward market colour, but they should be treated carefully and should not dominate the primary signal.
The ETA window determines whether inventory pressure is measured using only prompt and forward arrivals, or whether backward-looking confirmation windows are also included.
| Rank | ETA Window | Best Spearman | P90 Spearman | Top 100 Share |
|---|---|---|---|---|
| 1 | 0 to 4 | 0.6155 | 0.2234 | 100% |
| 2 | -6 to 4 | 0.5405 | 0.2002 | 0% |
The 0 to 4 week ETA window dominated.
It appeared in 100% of the top 100 configurations and produced the strongest observed relationship.
The narrower forward-looking window consistently outperformed the broader window that incorporated six weeks of backward-looking arrival confirmation. This suggests that the strongest pricing relationships were associated with prompt and forward observable supply pressure rather than inventory definitions that incorporated older arrival information.
The primary framework should focus on the current week through four weeks forward.
Backward confirmation may still have research value, but the strongest pricing relationships emerged from prompt and forward observable supply pressure.
The framework tested named weighting methodologies to determine whether prompt, forward, balanced, or backward-confirmation weighting curves improved signal construction.
These weighting sets were designed to test whether different arrival timing assumptions improved the relationship between inventory pressure and pricing.
| Rank | Weighting Set | Best Spearman | P90 Spearman | Top 100 Share |
|---|---|---|---|---|
| 1 | target_week_only | 0.5405 | 0.2469 | 0% |
| 2 | backward_confirmation | 0.3460 | 0.2197 | 0% |
| 3 | tight_prompt | 0.3898 | 0.2055 | 0% |
| 4 | balanced_curve | 0.4303 | 0.1924 | 0% |
| 5 | forward_heavy | 0.4298 | 0.1931 | 0% |
The named weighting sets produced useful results, but they did not appear in the top 100 strongest configurations.
This does not mean weighting is irrelevant. Rather, the strongest models came from rolling and decayed inventory accumulation signals that did not rely on these named weighting_set_name fields.
The best named weighting result came from target_week_only, which suggests that when simple weighting systems are used, prompt and target-week supply are more informative than more distributed weighting curves.
Named weighting sets should remain part of the research framework, but the primary signal architecture should prioritise rolling and decayed accumulation systems.
Prompt weighting remains economically intuitive, but persistence-based inventory construction performed better.
Signal construction was one of the most important research questions.
The framework tested whether pricing relationships are stronger when inventory pressure is measured as:
| Signal Family | Purpose |
|---|---|
| Volume level | Absolute inventory pressure. |
| Volume change | Inventory acceleration or deceleration. |
| Z-score | Abnormal inventory pressure versus history. |
| Binary high inventory | High-pressure regime indicator. |
| Rolling sum | Persistent accumulated inventory pressure. |
| Decayed inventory | Time-weighted accumulated inventory pressure. |
| Signal Type | Best Spearman | P90 Spearman | Top 100 Share |
|---|---|---|---|
| decayed_equal_6w_inventory | 0.6155 | 0.5910 | 15% |
| rolling_6w_inventory_sum | 0.6154 | 0.5909 | 15% |
| rolling_8w_inventory_sum | 0.6072 | 0.5829 | 10% |
| decayed_equal_8w_inventory | 0.6072 | 0.5829 | 10% |
| decayed_equal_4w_inventory | 0.6053 | 0.5831 | 12% |
| rolling_4w_inventory_sum | 0.6053 | 0.5831 | 12% |
| rolling_3w_inventory_sum | 0.5963 | 0.5819 | 8% |
| decayed_linear_4w_inventory | 0.5876 | 0.5620 | 6% |
| decayed_front_loaded_4w_inventory | 0.5876 | 0.5620 | 6% |
| rolling_2w_inventory_sum | 0.5638 | 0.5611 | 0% |
| volume_zscore | 0.5405 | 0.2373 | 0% |
| volume_level | 0.5386 | 0.2559 | 0% |
| binary_high_inventory | 0.3184 | 0.1757 | 0% |
| volume_change | 0.1831 | 0.1129 | 0% |
Signal type was the most important design factor in the entire framework.
Rolling and decayed inventory accumulation systems clearly outperformed simple weekly volume level, z-score, binary, and change signals.
The strongest observed signal was decayed_equal_6w_inventory, with a best Spearman of 0.6155. rolling_6w_inventory_sum was effectively identical, with a best Spearman of 0.6154.
The top 100 models were heavily concentrated in rolling and decayed accumulation systems:
| Signal Type | Top 100 Share |
|---|---|
| decayed_equal_6w_inventory | 15% |
| rolling_6w_inventory_sum | 15% |
| decayed_equal_4w_inventory | 12% |
| rolling_4w_inventory_sum | 12% |
| decayed_equal_8w_inventory | 10% |
| rolling_8w_inventory_sum | 10% |
The strongest results consistently emerged from four-to-eight-week inventory accumulation systems. This strongly supports the conclusion that persistent inventory pressure contains substantially more information than isolated weekly inventory observations.
The primary signal architecture should focus on rolling and decayed inventory accumulation, particularly four-to-six-week systems.
Simple week-on-week inventory change was not a strong signal in this framework.
Rolling windows test whether sustained inventory pressure is more important than a single weekly observation.
The framework tested two, three, four, six, and eight-week rolling periods.
| Rolling Window | Best Spearman | P90 Spearman | Top 100 Share |
|---|---|---|---|
| 2 weeks | 0.5638 | 0.5611 | 0% |
| 3 weeks | 0.5963 | 0.5819 | 8% |
| 4 weeks | 0.6053 | 0.5689 | 36% |
| 6 weeks | 0.6155 | 0.5660 | 36% |
| 8 weeks | 0.6072 | 0.5589 | 20% |
The best single result came from the six-week window, while the top 100 models were dominated by four and six-week windows.
Together, four and six-week rolling windows accounted for 72% of the top 100 model appearances.
The results suggest that inventory pressure becomes more informative when accumulated across multiple weeks rather than measured as a single-period observation. Performance improved materially between the shorter two and three-week windows and the longer four, six, and eight-week windows.
This supports the conclusion that inventory pressure operates as a persistent market-state variable rather than a purely short-term event signal.
Four-to-six-week rolling inventory systems should be prioritised in future research and monitoring.
Eight-week windows remain useful, but may begin to dilute the immediacy of the physical signal.
Decay type tests whether recent observations should receive more weight than older observations when constructing accumulated inventory signals.
The framework tested equal, linear, exponential, and front-loaded decay structures.
| Decay Type | Best Spearman | P90 Spearman | Top 100 Share |
|---|---|---|---|
| equal | 0.6155 | 0.5862 | 37% |
| linear | 0.5911 | 0.5619 | 12% |
| front_loaded | 0.5876 | 0.5475 | 6% |
| exponential | 0.5821 | 0.5485 | 0% |
Equal weighting performed best.
It produced the strongest observed Spearman relationship, the highest P90 Spearman value, and the largest share of top 100 model appearances.
This is a notable result. It suggests that, within the tested rolling windows, sustained inventory pressure matters more than aggressively prioritising the most recent week.
In other words, the market relationship appears to be driven by persistent supply pressure accumulated across several weeks rather than only the freshest arrival information.
The fact that more aggressive decay structures failed to outperform equal weighting suggests that inventory persistence may be more important than recency alone when explaining future Singapore HSFO crack behaviour.
Equal-weight rolling and decayed accumulation systems should remain the default for the strongest inventory pressure framework.
More complex decay functions may be useful for future modelling, but they did not outperform equal accumulation in this research set.
The framework tested whether inventory pressure relates more strongly to crack levels, crack changes, standardised crack levels, or standardised crack changes.
| Rank | Target Type | Best Spearman | P90 Spearman | Top 100 Share |
|---|---|---|---|---|
| 1 | crack_level | 0.6155 | 0.4190 | 100% |
| 2 | crack_zscore | 0.5405 | 0.1656 | 0% |
| 3 | crack_change | 0.1915 | 0.1110 | 0% |
| 4 | crack_change_zscore | 0.1831 | 0.1035 | 0% |
The relationship was strongest against crack levels, not crack changes.
This is an important structural finding. It suggests that observable inventory pressure is more closely associated with the state of the forward crack structure than with short-term weekly price movements.
The results also indicate that level-based targets outperformed both change-based and standardised target definitions. The strongest relationships emerged when inventory pressure was compared directly against the underlying crack structure itself.
That supports the interpretation of inventory pressure as a market-state variable rather than a short-term directional trigger.
The framework should continue to prioritise crack level targets for structural research.
Crack change targets may be more relevant for future trading signal development, but they were not the strongest target type in this phase.
The framework tested whether inventory pressure relates to future pricing behaviour across one-to-four-week horizons.
The strongest results would indicate where physical supply pressure appears to transmit into pricing structure.
| Target Feature | Best Spearman | P90 Spearman | Top 100 Share |
|---|---|---|---|
| target_crack_lag_3 | 0.5953 | 0.5245 | 26% |
| target_crack_lag_2 | 0.6155 | 0.5376 | 74% |
| target_crack_lag_4 | 0.5807 | 0.5176 | 0% |
| target_crack_lag_1 | 0.5121 | 0.2953 | 0% |
| Lag Weeks | Best Spearman | P90 Spearman | Top 100 Share |
|---|---|---|---|
| 3 | 0.5953 | 0.2071 | 26% |
| 4 | 0.5807 | 0.2131 | 0% |
| 2 | 0.6155 | 0.2197 | 74% |
| 1 | 0.5405 | 0.1862 | 0% |
The strongest single results came from the two-week forward crack level target.
The top 100 results were concentrated entirely in two and three-week forward crack level targets, with the two-week horizon appearing in 74% of the top 100 models.
This supports the idea that physical inventory pressure does not always transmit instantly into benchmark pricing. A two-to-three-week window appears to be the strongest observed pricing transmission horizon.
The two-week forward crack level should be treated as the primary target horizon for the current framework.
The three-week horizon is also important and should remain part of monitoring and future development.
The top 100 model frequency analysis provides the clearest view of recurring structure.
| Factor | Dominant Top-100 Result |
|---|---|
| Confidence | confidence_clean appeared in 100% |
| Product | hsfo_only appeared in 100% |
| ETA Window | 0 to 4 appeared in 100% |
| Target Type | crack_level appeared in 100% |
| Lag | lag 2 appeared in 74%, lag 3 in 26% |
| Stem Status | active_stem_only and confirmed_only each appeared in 43% |
| Geography | extended_bunker_supply_region appeared in 37%, core_singapore_market in 35%, broad_south_east_asia in 28% |
| Signal Type | rolling and equal-decayed inventory accumulation dominated |
| Rolling Window | 4-week and 6-week windows each appeared in 36% |
| Decay Type | equal decay appeared most frequently |
The same pattern becomes visible when the strongest 100 configurations are analysed collectively. Several framework characteristics repeatedly appeared among the strongest-performing inventory systems.
Most notably, confidence_clean, hsfo_only, the 0 to 4 ETA window, and crack_level targets appeared in 100% of the top 100 configurations. This level of concentration is unusually strong and indicates that the strongest results were clustered around a highly consistent set of design choices.
The most important conclusion is that the strongest models were not scattered randomly across the experimental universe.
They clustered around a coherent structure:
This consistency materially increases confidence in the economic interpretation of the results.
The strongest observed configuration was:
| Factor | Level |
|---|---|
| Geography | extended_bunker_supply_region |
| Product | hsfo_only |
| Stem Status | active_stem_only |
| Confidence | confidence_clean |
| ETA Window | 0 to 4 |
| Signal Type | decayed_equal_6w_inventory |
| Rolling Window | 6 |
| Decay Type | equal |
| Target Type | crack_level |
| Target Feature | target_crack_lag_2 |
| Lag Weeks | 2 |
This configuration produced:
| Metric | Result |
|---|---|
| Spearman | -0.6155 |
| Pearson | -0.5212 |
| Observation Count | 202 |
| Combined Score | 0.5684 |
The direction of the relationship was negative.
The strongest observed configuration is illustrated below. The relationship demonstrates how sustained observable HSFO inventory pressure was associated with weaker future Singapore HSFO crack levels across the research period.
Higher observable HSFO inventory pressure was associated with weaker future Singapore HSFO crack levels.
The strongest observed model is not the same as the best average factor-level profile.
The strongest observed model identifies the single highest-performing configuration across the entire experimental universe. By contrast, the best average factor-level profile identifies which factor levels generally performed best within their own factor group when assessed across all experiments.
| Factor | Best Average Level |
|---|---|
| Confidence | confidence_clean |
| Geography | core_singapore_market |
| Product | hsfo_only |
| Stem Status | active_stem_only and confirmed_only |
| ETA Window | 0 to 4 |
| Signal Type | rolling_3w_inventory_sum |
| Target Type | crack_level |
| Lag Weeks | 3 |
| Rolling Window | 3 |
| Decay Type | equal |
| Weighting Set | target_week_only |
This table should not be interpreted as a single model specification. Instead, it identifies the strongest-performing factor level within each individual factor category when evaluated on a factor-level basis.
As a result, some entries differ from the strongest observed configuration. This is expected. The profile reflects average factor-level performance rather than the highest-performing combination of factors.
The strongest actual observed model remains the decayed_equal_6w_inventory configuration against two-week forward crack levels.
The research supports several broader conclusions.
The strongest configurations demonstrate a meaningful relationship between observable HSFO inventory pressure and future Singapore HSFO crack structure.
The best observed Spearman of -0.6155 and Pearson of -0.5212 are strong for a physical commodity market framework built from observable cargo-level data.
Rolling and decayed inventory accumulation systems dominated the strongest results.
This suggests that persistent inventory pressure is more informative than a single weekly cargo snapshot.
HSFO-only appeared in 100% of the top 100 models.
This indicates that direct product relevance matters. Broader HSFO-adjacent product groupings did not improve the strongest signal.
confidence_clean appeared in 100% of the top 100 models.
This is one of the clearest findings in the study. Higher-quality cargo filtering materially improved the framework.
Active and confirmed cargoes dominated the strongest stem status results.
This suggests that the market relationship strengthens when inventory pressure is built from physically mature cargoes rather than lower-certainty projections alone.
The framework performed much better against future crack levels than against weekly crack changes.
This supports the interpretation that inventory pressure explains structural market state rather than short-term price noise.
The strongest results appeared at the two-week forward horizon, with three-week targets also represented in the top 100.
This suggests that observable physical pressure may take time to transmit through regional fuel oil balances and benchmark pricing.
The framework may be relevant for:
| Participant | Potential Use |
|---|---|
| Commodity hedge funds | Physical-to-paper market structure research. |
| Fuel oil trading desks | Inventory regime monitoring and forward crack context. |
| Bunker market participants | Regional supply pressure assessment. |
| Refinery optimisation teams | Fuel oil balance and margin context. |
| Market analysts | Physical intelligence integration into pricing frameworks. |
| Risk teams | Inventory-state awareness and market regime interpretation. |
The framework should not be viewed as a standalone trading signal.
Its value lies in structural interpretation. It helps explain whether observable physical supply conditions are consistent with stronger or weaker forward pricing structure.
The research is intentionally correlation-based.
Correlation does not prove causation.
The framework measures observable cargo-level inventory pressure, not complete tank inventory.
Some physical flows may be missed, misclassified, delayed, rerouted, blended, or commercially reallocated.
The pricing relationship may change across market regimes.
The research period may include specific structural conditions that do not repeat exactly in future periods.
Some experimental configurations were intentionally broad or exploratory and were not expected to perform strongly.
The results should therefore be interpreted as evidence of structural relationship, not as a deterministic forecasting model.
Future development should focus on several areas.
The framework should be tested across distinct market regimes, including tight markets, oversupplied markets, high-volatility periods, and low-volatility periods.
The strongest relationships should be assessed through time to determine whether they remain stable or cluster around specific subperiods.
Because Spearman relationships were often stronger than Pearson relationships, future work should explore non-linear and regime-based models.
Future modelling may combine inventory pressure with other physical and market variables, including bunker demand, refinery margins, arbitrage economics, freight, and forward curve structure.
The current framework is a structural research engine. A future phase could convert the strongest configurations into a monitored prediction or regime-classification system.
The framework could be extended to LSFO, VLSFO, SRFO, VGO, and other residual or blending-related product groups.
The Maven Fuel Oil Inventory Pressure Research Framework demonstrates that observable physical fuel oil inventory pressure exhibits meaningful and economically coherent relationships with future Singapore HSFO crack pricing behaviour.
The strongest observed configuration produced a Spearman correlation of 0.6155 and a Pearson correlation of 0.5212.
The strongest configurations were consistent, not random. They repeatedly relied on HSFO-only cargoes, high-confidence filtering, active or confirmed stems, prompt-to-forward ETA windows, rolling or decayed inventory accumulation, and two-week forward crack level targets.
Four framework characteristics appeared in 100% of the strongest 100 configurations: confidence_clean, hsfo_only, the 0–4 week ETA window, and crack_level targets. This concentration provides additional evidence that the strongest results were clustered around a highly coherent and repeatable market structure.
The framework was designed to answer five core research questions.
Does observable inventory pressure matter?
Yes. The strongest configurations consistently demonstrated meaningful inverse relationships between observable HSFO inventory pressure and future Singapore HSFO crack behaviour.
Which product definition works best?
HSFO-only cargo definitions performed best. HSFO-only appeared in 100% of the strongest 100 configurations and materially outperformed broader HSFO-related product groupings.
Which cargo certainty level matters most?
High-confidence and operationally mature cargoes performed best. Confidence-clean filtering appeared in 100% of the strongest 100 configurations, while active-stem and confirmed-only frameworks consistently outperformed broader cargo sets.
Which timing window matters most?
Prompt and near-forward arrivals performed best. The 0–4 week ETA window appeared in 100% of the strongest 100 configurations and consistently outperformed broader timing definitions.
Which signal construction method works best?
Rolling and decayed inventory accumulation systems produced the strongest relationships. Persistent inventory pressure proved more informative than isolated weekly cargo movements.
The central conclusion is that observable HSFO inventory pressure behaves less like a short-term event signal and more like a persistent physical market-state variable.
Higher observable inventory pressure was associated with weaker future Singapore HSFO crack levels.
Lower observable inventory pressure was associated with stronger future crack structure.
The research supports the institutional case for incorporating systematic physical cargo intelligence into fuel oil market analysis. It also provides a repeatable framework for monitoring regional inventory pressure, assessing forward crack structure, and developing future physical-to-paper pricing models.
Full experimental outputs, factor definitions, and statistical result sets are available for institutional review.
For additional information, institutional discussions, collaboration enquiries, or trial access to Maven Knowledge physical fuel oil intelligence datasets and reporting, please contact Maven Knowledge.
