Weekly macro-commodity note with scored probabilistic forecasts

C4 · Weekly macro-commodity note with scored probabilistic forecasts#

Objective#

Write a short weekly note on four markets that matter to an Indian investor: gold, crude oil, USDINR and the Nifty 50. Each note states explicit probabilities for well-defined events that resolve within one to four weeks, explains the reasoning through the transmission channels of M10–M12, and is scored when the events resolve. Over the capstone you build a track record that tells you, with numbers, whether your macro judgement is calibrated and whether it has resolution. The purpose is to replace narrative confidence with a scored forecast journal (M13 · 13.8). No note may contain a trade recommendation.

Deliverables#

1. Weekly notes (800–1,500 words each), published to your journal before the first resolution window opens, with a timestamp. At least four consecutive weeks; more is better.
2. Forecast register (CSV or spreadsheet): one row per forecast with id, date issued, question text, resolution rule, data source, horizon, probability, and after resolution the outcome (0/1) and resolution date. Probabilities are never edited after issue; updates are new rows that reference the old id.
3. Scoring report at the end: Brier score, log score, Murphy decomposition, calibration table and reliability diagram, Brier skill scores against two references, and per-market breakdowns.
4. Review essay (1,500–2,500 words): what you learned about your forecasting, with examples of your best and worst calls and the process changes you made.

Time budget#

About 1–1.5 hours a week for each note and the register (the study plan puts it on every review day once C4 starts), plus about 2 hours to set up the register and the first questions and 3–4 hours for the scoring report and the review essay at the end: about 12–15 hours over the six weeks the full-track plan runs it, and about 1–1.5 hours for each further week you continue. Twenty resolved forecasts (acceptance criterion 1) need at least four weeks of notes; calibration bins need more, so start C4 as early as the plan allows and keep it running after the course.

Required content of each weekly note#

1. State of the world (M11 · 11.15; M10 · 10.6, 10.8): levels and weekly changes of gold (USD and MCX INR), Brent and WTI, USDINR (FBIL reference rate), Nifty 50, US 10-year nominal and real yields, DXY, India 10-year G-sec, India VIX; each with source and date.
2. Calendar: scheduled events in the window (FOMC, RBI MPC, US CPI and payrolls, India CPI, OPEC+ meetings, EIA weekly reports, index expiries, elections or geopolitical deadlines).
3. Transmission analysis (M11 · 11.4, 11.11–11.12; M12 · 12.7): for the main live story, the chain from shock to each of the four markets, with signs and rough elasticities (for example the effect of a $10 move in Brent on India's current account and the rupee), citing where the elasticity comes from.
4. Quantitative inputs, at least two per note:
   • the market-implied distribution: forward premium on USDINR from CIP (M11 · 11.8), the implied volatility of Nifty options (M09) or MCX gold/crude options converted into an implied probability of a move of a given size (Black-76 risk-neutral probability $N(d_2)$, stated as risk-neutral, not real-world);
   • carry and roll yield on the gold and crude curves (M10 · 10.3);
   • positioning: CFTC managed-money net positioning and its z-score (M10 · 10.12), NSE participant-wise OI (M09A);
   • seasonality only if tested properly (M10 · 10.13);
   • scenario trees with Bayesian updating for geopolitical events (M12 · 12.5).
5. Forecasts: at least five new probabilistic forecasts per week, covering all four markets over the capstone, each with a one-paragraph rationale that names the base rate used.
6. Review of resolved forecasts from earlier notes, with a sentence on each miss.

Writing forecast questions#

Each question must be resolvable by a stated data source on a stated date with no judgement. Good forms:

• "Will MCX GOLD (near-month) settle above ₹X on date D?" (resolution: MCX bhavcopy settlement price).
• "Will Brent front-month settle below $Y on any day before D?" (resolution: ICE settlement or EIA daily Brent spot, stated in advance).
• "Will the FBIL USDINR reference rate on D be above Z?"
• "Will the Nifty 50 close on D be within ±2% of its close on the issue date?"
• "Will the RBI change the repo rate at the MPC meeting ending on D?"

Set thresholds so that your probabilities are spread over the range: forecasting only near-certain events produces a good Brier score and no information. Use the market-implied probability as the reference and record whether you are above or below it, and why.

Data sources#

• Gold: Yahoo GC=F; MCX bhavcopy (manual download; mcxindia.com blocks automated fetching); IBJA rates; World Gold Council data.
• Crude: EIA daily Brent and WTI spot (FRED DCOILBRENTEU, DCOILWTICO); EIA Weekly Petroleum Status Report; OPEC press releases.
• USDINR: FBIL reference rate; RBI DBIE; FRED DEXINUS (noon buying rate, lagged); NSE currency derivatives bhavcopy for futures prices.
• Nifty: NSE index archives; Yahoo ^NSEI; NSE F&O bhavcopy for options; India VIX.
• Rates and dollar: FRED DGS10, DFII10, DTWEXBGS; RBI for G-sec yields and policy rates.
• Positioning: CFTC Commitments of Traders, disaggregated futures-only history files (cftc.gov); NSE participant-wise OI.
• Geopolitics: GPR and AI-GPR indices (Caldara–Iacoviello; M12 · 12.4).

Acceptance criteria#

1. At least 20 resolved forecasts over at least four weeks, every one issued before its resolution window opened, with no edits after issue (the register's timestamps show it).
2. Every forecast has a resolution rule and data source written at issue; disputed resolutions are resolved by the rule, not by judgement after the fact.
3. Probabilities span at least four of the five bins 0–0.2, 0.2–0.4, 0.4–0.6, 0.6–0.8, 0.8–1.0.
4. The scoring report contains:
   • the Brier score and the log score (M13 · 13.7), with the log score computed with probabilities clipped to [0.01, 0.99] and the clipping stated;
   • the Murphy decomposition $\mathrm{BS} = \mathrm{REL} - \mathrm{RES} + \mathrm{UNC}$ (M13 Proposition 13.20) on the five bins above, verified to sum to the Brier score computed directly;
   • a calibration table: for each bin, number of forecasts, mean forecast probability, observed frequency, and a binomial 90% interval for the observed frequency;
   • Brier skill scores against the coin flip (0.5) and against climatology (the overall base rate), and, where available, against the market-implied probability;
   • a breakdown by market and by horizon.
5. The review essay identifies at least one systematic bias in your forecasts (overconfidence, a directional lean, neglect of base rates) with the evidence from the table, and states a concrete process change.
6. All changeable facts in each note carry an as-of date and source.

Template outline for each weekly note#

1. Headline numbers table (with dates and sources).
2. The week's main story and its transmission to gold, crude, USDINR and the Nifty.
3. Quantitative inputs: implied distributions, carry, positioning, scenario tree.
4. Calendar for the coming window.
5. New forecasts (table: id, question, resolution rule, horizon, probability, market-implied probability, rationale).
6. Resolved forecasts since the last note and short commentary.
7. Running scores (Brier, log, count).

Common failure modes#

• Ambiguous questions ("Will gold be strong?") that cannot be scored. Every question needs a number, a date and a source.
• Editing probabilities after the news. Updates are new forecasts; the original stays in the record.
• Only easy questions. A record of 95% forecasts on near-certain events has a low Brier score and no resolution. The Murphy decomposition exposes this: RES near zero.
• Confusing risk-neutral with real-world probabilities. An option-implied probability includes risk premia; for crash events it overstates the real-world probability (M09 · 9.4). Say which you are quoting.
• Narrative without base rates. Start from how often a move of that size happens over that horizon (from the historical distribution or the implied volatility), then adjust.
• Too few forecasts per bin to say anything about calibration; widen bins or issue more forecasts rather than over-interpret five data points (M13 · 13.7, Worked example 1).
• Recommendations creeping in. The note forecasts events; it does not tell readers to buy or sell.