Forecast Scorecard
Running accuracy log of every forecast written via the predictive query workflow.
Aggregate Stats
| Metric | Value |
|---|---|
| Total forecasts written | 2 (3 sub-components: Turkey K, Turkey FM, Pakistan) |
| Total fully resolved | 2 (Turkey Shredded + Pakistan; all 3 sub-components) |
| Pending (no actual available) | 0 |
| Hit rate (in range) | 3/3 sub-components = 100% |
| Direction accuracy | 3/3 = 100% |
| Mean absolute error vs central | $0.98/t (K $0.50, FM $0.64, Pakistan $1.79) |
| Median absolute error | $0.64/t |
| Best forecast (smallest error) | Turkey Shredded K — $0.50/t (+0.12%) |
| Worst forecast (largest error) | Pakistan CFR Port Qasim — $1.79/t (−0.42%) |
Note: errors remain small (all < $2/t). The two Turkey actuals landed slightly above central (Eid slowed signal propagation); Pakistan landed below central as the bearish ex-UK Turkey HMS signal transmitted fully, pulling it to the low end of the range.
By Market
Turkey Shredded Scrap CFR
| K Component | FM Component | |
|---|---|---|
| Forecasts | 1 | 1 |
| Resolved | 1 | 1 |
| Hit rate | 100% | 100% |
| MAE | $0.50/t | $0.64/t |
| Direction accuracy | 100% | 100% |
| Range used | $426–$430 | $418–$425 |
| Actual | $428.5 | $421.64 |
Pakistan Scrap Import (CFR Port Qasim)
| MB-STE-0887 | |
|---|---|
| Forecasts | 1 |
| Resolved | 1 |
| Hit rate | 100% |
| MAE | $1.79/t |
| Direction accuracy | 100% |
| Range used | $420–$426 |
| Actual | $420.21 (05 Jun assessment) |
Recent Resolutions
| Target date | Market | Sub-bench | Forecast range | Actual | Hit | Error $/t |
|---|---|---|---|---|---|---|
| 2026-06-01 | Turkey shredded | K (Kallanish) | $426–$430 (c: $428) | $428.5 (Jun 4) | ✓ | +$0.50 |
| 2026-06-01 | Turkey shredded | FM MB-STE-0095 | $418–$425 (c: $421) | $421.64 (Jun 3) | ✓ | +$0.64 |
| 2026-06-01 | Pakistan CFR PQ | FM MB-STE-0887 | $420–$426 (c: $422) | $420.21 (Jun 5) | ✓ | −$1.79 |
Methodology Notes
Scoring window: Forecasts are scored against the FIRST available actual print ON OR AFTER the target date. If the target falls during a market holiday (Eid, Christmas, etc.) and the index print is delayed, scoring uses the first available post-holiday print. If no print appears within 10 business days of target, the forecast is marked “lapsed” and scored against the last pre-target print with a penalty flag.
Hit / Partial / Miss:
- Hit (✓): Actual falls within the forecast range.
- Partial (~): Actual is within $1/t of either range boundary (one step outside).
- Miss (✗): Actual is more than $1/t outside the range.
Error: Signed error = actual − central tendency. Positive = above central; negative = below.
Sources: Actuals sourced from entity pages in wiki/entities/ which are updated weekly from raw/extracted/ publisher files. Publisher = Kallanish (K) for daily consensus; Fastmarkets (FM) for transaction-index benchmarks.
Learning Log
Turkey Shredded Jun 1 — What Worked
The K shredded–K HMS $20/t mechanical spread was the most reliable driver. The forecast built this in explicitly and it held exactly ($404.5 HMS → $428.5 K shredded = $24, close to historical $20–21). The Rotterdam FOB shredded anchor ($388 + $34 freight = $422 implied) correctly predicted FM’s floor. The main miss in the reasoning: the Eid holiday slowed signal propagation by ~5 days, so actuals landed at the upper-middle of ranges (high $428.5 K vs $426 low-end) rather than the lower middle. Range width was appropriate; timing assumption was slightly off.
Learning: Build a 5-day Eid/holiday timing buffer into the horizon when a major market holiday falls between forecast generation and target date. The ex-UK deal signal is reliable but takes longer to transmit when the main buyer (Turkey) is on holiday.
Pakistan Jun 1 — What Worked
The Pakistan = Turkey-K-shredded − $5/$6 spread model resolved cleanly: as the ex-UK $406 Turkey HMS deal transmitted (HMS → $404.5 → $398.5), Pakistan followed to $420.21, a HIT inside the $420–$426 range. The forecast correctly weighted the bearish K-spread chain over the FM-anchoring case, but the central tendency ($422) was still ~$1.8 high — the actual landed at the very bottom of the range, exactly where the “downside scenario” ($419–421) pointed.
Learning: When the bearish driver is a confirmed transacted deal (not just an assessment), lean the central tendency toward the bearish-scenario band rather than splitting the difference with the anchoring case. The $0.21 in-range margin shows the range was correctly placed but the point estimate under-weighted the strength of a confirmed-deal signal.