On April 9, 2025, gold jumped 2.97% in a single day. Trump had announced sweeping tariffs. Every outlet ran it as breaking news. The next day, gold surged another 3.23% when he announced a 90-day pause. Both moves were genuinely rare — that magnitude happens less than once a month. Three days earlier, on April 6, gold had moved 0.3%. A headline still ran: "Gold edges higher on geopolitical uncertainty." That move happens on two out of every three trading days. It was noise dressed as signal. How different are these two cases, really? The first describes an event where less than 2% of trading days produce a move that large. The second describes something more common than a coin flip. One is worth understanding. The other exists because a journalist needed to file copy. This distinction — whether a market move is statistically meaningful or just random fluctuation — is something you can determine with a single division, once you understand the framework behind it. This post builds that framework from scratch: what it means for a move to be "large," how to measure it, and where the method breaks down. --- ## The problem ### A price on its own tells you nothing Gold closed at $2,658 on January 2, 2025. Was that a good day or a bad day? Was the move large or small? Was it driven by news, or was it random drift? You can't answer any of these questions from the price alone. A price is a point with no context. What matters is not where a price *is* — it's how much it *moved*, and whether that move was large or small relative to what normally happens. Financial news writes headlines about prices every day. "Gold closes at $3,200." "Nifty at 24,500." These numbers feel meaningful. They aren't. ### The question that actually matters Given how this asset normally behaves, was today's move routine or unusual? Two parts. First: what is the typical daily volatility for this asset? Second: how does today's move compare to that baseline? Edward Thorp — mathematician, hedge fund manager, author of *Beat the Dealer* and *A Man for All Markets* — could answer this in his head in seconds. He'd hear a news commentator predict the Dow would fall 9 points and know instantly: noise. We'll work through his exact reasoning later. First, the building blocks. --- ## The shape of normal ### Daily returns — the universal language Stop thinking in prices. Think in percentage changes. ``` return = (today's price − yesterday's price) ÷ yesterday's price × 100 ``` A 1% move means the same thing whether gold is at $1,000 or $3,000. Percentages are comparable across assets, across time, across price levels. Raw prices aren't. For 2025 gold, the first few returns: | Day | Calculation | Return | |-----|-------------|--------| | Jan 2 → Jan 3 | ($2,645 − $2,658.90) ÷ $2,658.90 | −0.52% | | Jan 3 → Jan 6 | ($2,638.40 − $2,645) ÷ $2,645 | −0.25% | | Jan 6 → Jan 7 | ($2,656.70 − $2,638.40) ÷ $2,638.40 | +0.69% | You now have 251 such numbers for 2025 — one for each trading day. The question becomes: given this collection, what counts as "normal" and what counts as "unusual"? ### Why the bell curve Take those 251 returns, sort them into buckets, and count how many fall in each. The tallest bars cluster near zero. The distribution falls away symmetrically on both sides. It looks like a bell. This isn't an assumption we impose. It's a pattern that emerges. Each daily return is the net result of dozens of independent forces — dollar moves, fund flows, central bank commentary, options expiries, geopolitical news. When many independent forces combine, the Central Limit Theorem says the result tends toward a bell shape. The bell isn't a theory we apply to gold. It's a consequence of how markets work. ### Standard deviation — the step size Every asset has its own bell. Some are wide (volatile), some are narrow (stable). One number determines the width: the standard deviation, written σ (sigma). σ answers the question: what is the typical daily move for this asset? For 2025 gold, σ = **1.29%**. For crude oil, σ = **1.98%**. For S&P 500, σ = **1.18%**. For the Nifty 50, σ = **0.75%**. Computing σ from data: ``` 1. Compute the mean of all returns (gold 2025: mean = +0.20%/day) 2. For each return, subtract the mean 3. Square each deviation 4. Average all the squared deviations → variance 5. Square root of the variance → σ Gold 2025: variance = 1.666 → σ = √1.666 = 1.29% ``` σ is the step size on the bell curve's x-axis. One step on gold's curve = 1.29%. Two steps = 2.58%. This "step" is the natural unit for measuring how far a move is from the center. ### The bell curve — visualized The x-axis is in σ-steps. The height represents how many days land at each point. Most cluster near zero; fewer and fewer land at the extremes. ![The bell curve with 68-95-99.7 sigma zones for gold 2025](/writing/signal-vs-noise/bell-curve-zones.png) ### The 68–95–99.7 rule The bell's fixed shape gives you three numbers worth knowing: - **68% of days** fall within ±1σ. Two in three days, gold moves less than ±1.29%. - **95% of days** fall within ±2σ. Only 5% of days exceed ±2.58%. - **99.7% of days** fall within ±3σ. Beyond ±3.87% is less than 1 in 300 days. Verified against real 2025 gold data: | Band | Expected | Actual (251 days) | |------|----------|-------------------| | Within ±1σ | 68.3% | 74.9% | | Within ±2σ | 95.4% | 94.0% | | Within ±3σ | 99.7% | 99.2% | Close enough. The slight excess in ±1σ and deficit in ±3σ reflects fatter tails in real financial data — a known property. For the purposes of filtering headlines, the approximation holds. --- ## The ruler ### Z-score — the x-axis, relabeled Look at the bell curve's x-axis. The tick marks at −3, −2, −1, 0, +1, +2, +3 — those ARE z-scores. The z-score is simply your position on the bell curve, expressed in units of σ. z = 1 means you're standing at the +1σ mark. z = 2 means +2σ. That's all it is. For gold (σ = 1.29%): z = 1 corresponds to a move of 1.29%. z = 2 corresponds to 2.58%. ### Why divide by sigma ``` z = move ÷ σ ``` This is a unit conversion. The move is in %. The bell curve's x-axis is in σ-units. Division by σ converts your move into "number of steps from the center." A 1.9% move in crude oil (σ = 1.98%) is z = 0.96 — about one step. Marginal. A 1.9% move in gold (σ = 1.29%) is z = 1.47 — one and a half steps. Same absolute move. Different meaning. The σ-context determines whether it's worth paying attention. ### The mental model A flat road. You stand at center (z = 0). The ground is marked at regular intervals. Each interval is one σ-step. ``` ← −3σ −2σ −1σ [YOU] +1σ +2σ +3σ → −3.87% −2.58% −1.29% 0 +1.29% +2.58% +3.87% one step = σ = 1.29% (gold 2025) ``` A headline says gold moved +0.4%. You compute z = 0.4 ÷ 1.29 = 0.31. Less than a third of one step. You barely shuffled off the origin. A headline says gold moved +2.58%. z = 2.58 ÷ 1.29 = 2.0. Two full steps from center. Now you're at the boundary of what 95% of days produce. ### Area = probability (and we want the OUTSIDE) Here's the critical insight that makes this useful. The bell curve's total area = 100%. Any region's area = the fraction of days landing there. The question we're asking is NOT "what fraction of days land inside my position?" We want the opposite: **"what fraction of days produce a move AT LEAST as large as mine?"** That's the area OUTSIDE — the two tails beyond your position. ![Noise vs Signal — the outside tails are what matter](/writing/signal-vs-noise/noise-vs-signal-tails.png) Why two tails? Because direction doesn't matter for this kind of filtering. A −2% drop is as statistically notable as a +2% gain. We shade BOTH sides beyond ±z. That combined outside area is the probability we care about. At z = 0.5 (left panel above), the red tails are enormous — most of the curve is outside your flags. The move is smaller than what happens on more than half of trading days. At z = 2.0 (right panel), the red tails are tiny slivers. Almost the entire curve is inside. The move exceeds what 95% of ordinary days produce. **The full probability table:** | z | Outside area (both tails) | Meaning | Verdict | |---|---|---|---| | 0 | 100% | Every day moves at least this much | — | | 0.5 | 62% | 3 out of 5 days are bigger | Noise | | 1.0 | 32% | 1 in 3 days are bigger | Marginal | | 1.5 | 13% | 1 in 8 days are bigger | Signal | | 2.0 | 4.6% | 1 in 22 days are bigger | Strong | | 2.5 | 1.2% | 1 in 80 days are bigger | Major | | 3.0 | 0.27% | 1 in 370 days are bigger | Extreme | If 62% of random days produce a bigger move than yours (z = 0.5), your move is unremarkable. If only 1.2% of days produce a bigger move (z = 2.5), something genuinely unusual happened. ![The probability ladder — z-score to verdict](/writing/signal-vs-noise/probability-ladder.png) --- ## The filter in practice ### Headline to verdict The complete process: ``` Step 1: Extract the move from the headline "Gold surges 2.1% on Fed pivot fears" → move = 2.1% Step 2: Recall σ for that asset Gold σ = 1.29% Step 3: Compute z z = 2.1 ÷ 1.29 = 1.63 Step 4: Locate z on the probability table z = 1.63 sits between 1.5 (1 in 8 days) and 2.0 (1 in 22 days) Step 5: Classify z > 1.5 → Signal. Worth reading. ``` Once σ is memorized, the only real-time work is a single division. The rest is a table lookup. ### Quick reference for gold (2025) | Threshold | Move | z | What to do | |-----------|------|---|--------| | Ignore | < 0.65% | < 0.5 | Skip the headline entirely | | Notice | > 1.94% | > 1.5 | Something real may have happened — read further | | Alarm | > 3.23% | > 2.5 | Something definitely happened — go find out what | --- ## Worked examples ### Thorp's DJIA — the original insight This framework comes from a passage in Edward Thorp's *A Man for All Markets*: > "I heard the reporter say, 'The Dow Jones Industrial Average is down 9 points to 11,075 on fears of a further interest rate rise.' I mentally estimated a typical daily DJIA change at about 0.6 percent or about sixty-six points. The reported move of 9 points was less than a seventh of this. The probability of at least a 9-point move was about 90 percent, so the market action was very quiet and hardly indicative of any fearful response. Simple math allowed me to separate hype from reality." Walk through it: 1. **Move** = 9 points. σ = 0.6% × 11,000 = 66 points. 2. **z** = 9 ÷ 66 = 0.14. 3. **Lookup**: z = 0.14 is far below 0.5. On 89% of ordinary days, the market moves more than this. 4. **Verdict**: Noise. The commentator manufactured drama about a move that barely registers on the bell curve. The reporter attributed causation ("on fears of...") to a move so small it required no explanation at all. ### 2025 gold — real events, real z-scores Gold went from $2,658 to $4,325 in 2025 — up 63%. The major moves clustered around three macro events: the Trump tariff shock, the US-China trade ceasefire, and the October record run. **The big days (z > 2.5):** | Date | Move | z | What happened | |------|------|---|---| | Oct 21 | −5.74% | 4.44 | Steepest single-day gold drop in over a decade. Nine-week rally ended: profit-taking, easing US-China tensions, stronger dollar, and post-Diwali drop in Indian demand all hit simultaneously. | | Dec 29 | −4.50% | 3.49 | Year-end profit-locking. Gold was up 63% for 2025. Funds closed positions before Dec 31. Calendar mechanics, not macro news. | | Apr 23 | −3.66% | 2.84 | Reversal after three weeks of tariff-driven gains. Profit-taking by large holders. | | Oct 20 | +3.50% | 2.71 | Gold broke $4,000 for the first time (Oct 8). Continued surge on geopolitical tension and central bank buying. | | May 12 | −3.46% | 2.68 | US and China announced 90-day tariff ceasefire. US cut tariffs from 145% to 30%. Fear evaporated; safe-haven demand collapsed. | | Apr 16 | +3.35% | 2.60 | Tariff pause excluded China. US-China escalation continued. Safe-haven demand pushed gold to another record. | | Oct 13 | +3.34% | 2.59 | Gold surge continued on dollar weakness and institutional buying. | | Apr 10 | +3.23% | 2.50 | Trump announced 90-day tariff pause — gold hit new record on safe-haven inflows. | **The noise days (z < 0.01):** | Date | Move | z | |------|------|---| | Sep 10 | +0.008% | 0.006 | | Jul 7 | −0.009% | 0.007 | | Nov 6 | −0.010% | 0.008 | Financial news wrote stories on all three of these days. "Gold edges higher as dollar softens." "Gold slips on Fed commentary." These moves are so deep inside the first σ-step that they are indistinguishable from random drift. There is no signal here — only a journalist meeting a deadline. **The pattern**: every z > 2.5 day in 2025 had a real, identifiable macro event behind it. Every z < 0.3 day had nothing behind it except a word count requirement. ### Across assets — same framework, different σ | Asset | σ (2025) | "Ignore" threshold | "Alarm" threshold | |-------|----------|-------------------|-------------------| | Gold | 1.29% | < 0.65% | > 3.23% | | Crude Oil | 1.98% | < 0.99% | > 4.94% | | S&P 500 | 1.18% | < 0.59% | > 2.96% | | Nifty 50 | 0.75% | < 0.37% | > 1.87% | | Natural Gas | 4.49% | < 2.25% | > 11.2% | | Silver | 2.36% | < 1.18% | > 5.91% | Notice what this implies: a 2% move in gold (z = 1.55) crosses the signal threshold. The same 2% move in crude oil (z = 1.01) is merely marginal. And a 2% move on the Nifty (z = 2.67) is a major event — something that happens fewer than once every 80 trading days. Same percentage. Completely different statistical meaning. The σ-context is everything. --- ## The limits The z-score filter is useful precisely because it's simple. But simplicity comes with boundaries. Three things worth being explicit about. ### It tells you significance, not direction A z > 2 move means something real drove the market. It says nothing about what happens next. May 12, 2025 was z = 2.68 — and gold *fell* 3.46%. Knowing the move was statistically significant doesn't tell you whether it will continue or reverse. A trade in either direction still needs its own thesis. Z is a filter for your information diet. It tells you which headlines are worth reading. It does not tell you what to do with the information once you've read it. ### Sigma is not a constant — it drifts This is the most important caveat. The entire framework rests on knowing σ, and σ is not fixed. It changes over time as market conditions shift. **Gold sigma across 11 years (2015–2025):** | Year | σ | Context | |------|---|---------| | 2017 | 0.65% | Calmest year — low volatility everywhere | | 2020 | 1.35% | COVID shock — highest gold volatility in the decade | | 2025 | 1.29% | Trump tariff shock — second highest | | 11-year average | 0.94% | | The range is 0.65% to 1.35% — a **2x variation** across the decade. Using 2025's σ (1.29%) on a 2017-style calm market would make you too relaxed about moves that were actually unusual in that regime. **It also shifts within a single year.** Rolling 60-day sigma for gold during 2025: | Period | 60-day σ | vs annual 1.29% | |--------|----------|-----------------| | March (pre-tariff) | 0.84% | −35% | | June (post-tariff chaos) | 1.67% | +29% | | August (calm summer) | 0.92% | −29% | | November (October reversal) | 1.53% | +18% | In March the market was calm — a 1.3% move would have been z = 1.55, crossing the signal threshold. By June, volatility had doubled — the same 1.3% move was z = 0.78, well into noise territory. **Crude oil is far worse.** Oil's annual sigma ranged from 1.58% (2017) to 21.66% (2020 — when prices went negative). A single year's sigma for oil can be off by an order of magnitude. ### What to do about sigma drift Three practical approaches, in increasing order of effort: 1. **Accept the imprecision.** Even if your σ estimate is 30% off, the framework still works at the extremes. A 5% gold move is z > 3 under any plausible σ — always worth reading. A 0.2% move is z < 0.3 under any σ — always noise. The uncertainty only bites at the boundary (z ≈ 1.0 to 1.5) where you're deciding between "marginal" and "signal." 2. **Recalibrate quarterly.** Recompute σ from the most recent 60–90 days of data every few months. Catches regime changes with minimal effort. 3. **Use a rolling window.** Compute σ from the trailing 60 days rather than a full year. More responsive to volatility shifts, but noisier because fewer data points feed the estimate. For casual headline filtering, option 1 is sufficient. Know your asset's approximate σ. When in doubt, use a slightly lower σ — it makes your filter more sensitive, and it's better to read one extra headline than to miss a real event. ### Fat tails — rare events are less rare than the bell predicts Real markets have fatter tails than a perfect bell curve. Events the bell assigns near-zero probability to — z > 4, "should" happen once in 30,000 days — actually occur every few years. October 21, 2025 (z = 4.44) is an example. This doesn't break the filtering framework. Z > 2.5 still means "go look." But it means you should not use z-scores for precise probability estimation ("the chance of this is 0.01%"). The bell curve systematically underestimates tail risk. For filtering headlines, the imprecision doesn't matter. For pricing options or sizing portfolio risk, use a distribution with fatter tails. ### From filter to profit To go from "this headline is worth reading" to making money, you need three things this framework does not provide: 1. **A thesis** about what happens next (fundamental or technical analysis). 2. **Position sizing** — how much to bet relative to your conviction and edge. 3. **An edge** — some information or analytical advantage not already priced in. Z-score is a prerequisite for all three. It tells you when to pay attention. It does not tell you what to do with that attention. --- ## Summary The framework reduces to this: memorize σ for the assets you follow. When you see a headline, divide the reported move by σ. The result — z — tells you where the move lands on the bell curve. Below 0.5: noise, skip it. Above 1.5: signal, read further. Above 2.5: something real happened, understand why. The method has limits. Sigma drifts over time, so your calibration needs occasional updating. The bell curve underestimates extreme events. And statistical significance tells you nothing about direction or what to trade. But as a filter for the daily flood of financial media — for separating the 10% of headlines that contain real information from the 90% that don't — it works. --- *Not investment advice. This is a framework for evaluating information quality — for distinguishing meaningful market moves from routine fluctuation. It is not a trading strategy. All data sourced from Yahoo Finance (GC=F, CL=F, NG=F, SI=F, ^GSPC, ^NSEI) for the 2025 calendar year. Sigma values are period-specific and will change as market conditions evolve.*