# How to use this file AI for Data and Documents — SHOT LEARNING DEMO Zero-shot, One-shot, and Few-shot prompting on Fed statements WHAT "SHOTS" MEAN A "shot" is a worked EXAMPLE you give the AI inside the prompt before asking it to do the real task. - ZERO-SHOT : no examples. You just describe the task and ask. - ONE-SHOT : you give ONE example (an input and its correct answer), then ask. - FEW-SHOT : you give SEVERAL examples (input + correct answer pairs), then ask. You are NOT retraining the model — you teach it IN the prompt, from the examples. THE TASK Read a Federal Reserve (FOMC) statement and predict the rate decision it announces: HIKE / HOLD / CUT. Then reveal the ACTUAL decision and check. THE DOCUMENTS (download from the Policy Documents page -> "Shot Learning set") FOMC1 = March 15, 2020 (actual: CUT to 0-0.25%) FOMC2 = June 15, 2022 (actual: HIKE by 0.75%) FOMC3 = September 18, 2024 (actual: CUT by 0.50%) FOMC4 = April 29, 2026 (actual: HOLD) Keep the "actual" answers hidden from the model until you check. HOW TO RUN Each prompt below is self-contained. Copy it, open a NEW chat at claude.ai, upload the document it names, paste the prompt, and compare the answer to the real outcome. --- # Zero-shot — no examples Upload FOMC3 (September 2024), then paste: You are reading a Federal Reserve FOMC policy statement. Based ONLY on the language in this statement, predict the interest-rate decision it is announcing. Answer with exactly one word: HIKE, HOLD, or CUT. Then give one sentence of reasoning and one short quote that drove your answer. CHECK: the actual September 2024 decision was a CUT (0.50%). With no examples, the model relies only on its general reading of the tone. --- # One-shot — one worked example, then a new case In a NEW chat, upload FOMC3 (September 2024), then paste this prompt. It contains ONE solved example before the real question: Task: read an FOMC statement and classify the rate decision as HIKE, HOLD, or CUT. EXAMPLE 1 Statement excerpt: "Inflation remains elevated... The Committee is highly attentive to inflation risks... decided to raise the target range for the federal funds rate." Correct answer: HIKE Now classify the statement I have uploaded. Answer with one word (HIKE/HOLD/CUT), then one sentence of reasoning and a short supporting quote. CHECK against the actual CUT. Did one example change the answer or the reasoning? --- # Few-shot — several examples, then a new case In a NEW chat, upload FOMC4 (April 2026), then paste this prompt. It gives THREE solved examples first, each pairing statement language with the real outcome: Task: read an FOMC statement and classify the rate decision as HIKE, HOLD, or CUT. Learn from these examples, then classify the new statement. EXAMPLE 1 (March 2020) Language: "The effects of the coronavirus will weigh on economic activity... the Committee decided to lower the target range for the federal funds rate to 0 to 1/4 percent." Answer: CUT EXAMPLE 2 (June 2022) Language: "Inflation remains elevated... highly attentive to inflation risks... decided to raise the target range." Answer: HIKE EXAMPLE 3 (September 2024) Language: "Inflation has made further progress toward the Committee's 2 percent objective... decided to lower the target range." Answer: CUT NEW STATEMENT: classify the statement I uploaded. Give one word (HIKE/HOLD/CUT), one sentence of reasoning, and a short supporting quote. CHECK against the actual HOLD (April 2026). With several labeled examples, the model has a clearer pattern of how FOMC wording maps to decisions. --- # Compare and discuss After running the three prompts above, fill in this table: Approach | Examples given | Predicted | Actual | Correct? Zero-shot | 0 | ____ | CUT | ____ One-shot | 1 | ____ | CUT | ____ Few-shot | 3 | ____ | HOLD | ____ Then ask Claude: - "Explain how the examples I gave changed your reasoning between zero-shot and few-shot." - "What kind of task benefits most from few-shot examples, and when is zero-shot already enough?" - "What is the risk if my few-shot examples are biased or all point one way?" --- # Learn from a past document + outcome, then predict a new one This is the "give the model an old document AND the real outcome, then ask it to predict a new document" idea. Run it as three messages in ONE chat: STEP 1 (teach) — upload FOMC2 (June 2022) and paste: Here is an OLD FOMC statement and the REAL outcome that followed it: the Fed HIKED by 0.75%. Study which phrases in this statement went with a hike. Do not answer anything yet. STEP 2 (predict) — upload FOMC4 (April 2026) and paste: Here is a NEW statement. Using only what you learned from the old statement-and-outcome pair above, predict the decision (HIKE/HOLD/CUT) with a short supporting quote. Do not use any outside knowledge of what actually happened. STEP 3 (check): Reveal the actual decision (HOLD) and ask Claude whether one labeled past case was enough, and why several examples (few-shot) usually help. --- # The big lesson - Zero-shot is fast and needs no examples, but it leans entirely on the model's general knowledge and can miss domain-specific cues. - One-shot and few-shot "teach in the prompt": a few labeled examples often sharpen the answer a lot, with no coding and no retraining. - Always VERIFY against a known outcome when you can. A confident prediction is not the same as a correct one. - Biased or one-sided examples bias the answer. Choose examples that fairly represent the range of real outcomes.