Understanding scenarios and forecasts

1. Getting started

Without registering, you can enter assets, import public wallet addresses or CSV files, run scenarios and save temporary records. Guest holdings and the latest ten records stay in this tab’s session storage, not a personal cloud database. Refreshing retains them; closing the tab or clearing browser data may erase them. Export JSON first if needed. Signing in does not migrate guest data. Signing in with your own ChatGPT account is optional and enables cloud storage and access across devices. Public address lookup needs no sign-in or wallet connection: the address is sent to the server and Zerion, and guest results stay in this tab. Anonymous lookups have frequency and total usage limits; retry later if limited. No seed phrase or transaction signature is requested. Signing out hides cloud data without deleting holdings. Demo data is not saved to your account.

Manual amounts are USD valuations; currencies are not converted automatically. Wallet lookup failures show an error. Successful lookups do not guarantee coverage of every protocol or token. Check sources, assets and liabilities.

2. How scenarios are generated

The homepage uses fixed illustrative valuations and shocks, independent of market feeds and not a current market assessment. Try demo continues with the same numbers. Selecting a scenario or changing intensity updates portfolio results immediately; no second apply action is needed. Missing shock data produces no result rather than retaining another scenario’s outcome.

Conditions match only with five available global signals, two supporting conditions, complete required data and no opposing evidence. Support, opposition, missing and neutral evidence are listed separately. Risk recovery additionally checks deteriorating unemployment. Only a clear leader is applied automatically; ties and insufficient evidence require a manual choice. Refreshing preserves the selection. Risk appetite is a trend assumption; the others are macro transmission assumptions. Thresholds are experimental and not historically calibrated.

Daily series are compared with approximately 30 calendar days earlier; monthly series use the previous observation. Transmission paths are proposed conditional hypotheses awaiting validation, not established causal relationships. Each card lists current evidence and invalidation conditions.

Shock magnitude = sample volatility of up to 60 recent daily log returns for the representative asset × √21 (securities) or √30 (crypto) × a scenario multiplier. At least 20 return observations are required; shocks are capped at ±80%. Multipliers and directions are fixed by scenario: for example, Treasuries may rise during a growth slowdown, and oil may rise under inflation. Mild, base and strong settings apply 0.5, 1 and 1.5 times the base shock, with the same ±80% cap. These settings test sensitivity; they are not probability intervals or expected returns. More matching conditions do not increase the shock magnitude. Individual stocks may use class proxies that omit company-specific risks. Expand a card to see each applied shock. These are not regression estimates, return forecasts or event probabilities.

Asset classification

A-shares, Hong Kong equities and US equities have separate mappings. Generic equity classifications no longer default to US equities. Uncertain names require confirmation; edit and save their classification in the holdings list. Auto-detection uses limited rules and is not a security-symbol validation service.

A-shares use ASHR’s USD price as a broad-market volatility proxy. It includes currency, trading-time and tracking effects and does not cover stock-specific risk. ASHR tracks the CSI 300; see the official fund prospectus. Missing A-share data is not replaced by US equity data.

3. How holdings are mapped

BTC → BTC-USD; ETH/WETH → ETH-USD; US equities → SPY; China/Hong Kong equities → FXI; global equities → ACWI; intermediate US Treasuries → IEF; gold → GLD; oil → CL=F. Each position’s mapping can be changed in advanced settings.

Only the representative asset directly matches its quote. Individual stocks, other tokens, other bond maturities and other commodities use class proxies and may behave differently. FXI does not represent all A-shares; GLD is not the gold price per ounce. Property and unidentified assets lack a model, remain unchanged by default and are marked uncovered. Other crypto assets default to a BTC proxy and require confirmation.

Net value after a scenario = Σ [USD asset value × (1 + asset shock) − USD liability value × (1 + liability shock)]. On-chain debt uses the corresponding asset factor by default; other debt uses the USD cash factor. Freezing a location reduces liquid value, not book value, and does not remove liabilities. Stablecoins remain unchanged by default; use custom shocks to test depegging risk.

4. Price forecasts and confidence

Price forecasts use a separate fixed statistical baseline; macro scenario shocks are not price targets. Coverage is limited to BTC, ETH, SPY, FXI, ACWI, IEF, GLD and continuous WTI futures. At least 120 valid daily prices are required and stale data is excluded. Quotes from the current UTC date are excluded to reduce incomplete-day effects.

For the latest 60 daily log returns, μ is the mean and σ the sample standard deviation. The central median forecast is current price × exp(0.25 × μ × h). The horizon h is 21 trading observations for securities or 30 daily observations for crypto. The trend shrinkage coefficient 0.25 is a fixed assumption, not the result of parameter optimization.

Nominal 80% prediction interval = exp[log(center) ± 1.28155 × σ × √(h × (1 + 0.25² × h / 60))]. It assumes approximately normal daily log returns and continuing recent volatility. Crises, jumps, correlations and abrupt policy changes may invalidate it.

Historical evaluation starts at the 90th price observation. Each origin uses only prior data, with origins spaced h observations apart so target windows do not overlap. Results show interval hits, coverage and central-forecast mean absolute percentage error against an unchanged-price baseline. Historical prices may be revised by providers; this is not a strictly archived daily trading backtest. It excludes costs and dividends, and continuous oil contracts are affected by rolls.

Confidence is not the probability of a price increase or an accuracy rate. Fewer than 20 tests means Insufficient samples. Medium requires at least 20 tests, 70–90% coverage, mean error no worse than the unchanged-price baseline and fresh data. Otherwise confidence is Low. These tiers are quality indicators, not calibrated probabilities; there is no High tier. The 80% level is a model assumption; actual historical coverage is shown separately.

Method references: Time-series cross-validation, Prediction interval. The coefficients, thresholds and scenario rules are specific to this test product.

5. Data quality

Retrieval time is when the API was queried; observation date is when the value applies. Monthly CPI year-on-year change is calculated from a seasonally adjusted index and can differ from common unadjusted measures. When BLS is rate-limited, FRED CPIAUCSL and UNRATE provide equivalent source series. The FRED key stays on the server.

Opening the page refreshes a cache older than 30 minutes. Manual refreshes are at least one minute apart; BLS monthly batch requests are attempted at most once per day. Observations are disabled after 5 days for securities, 3 for crypto, 10 for daily macro, 75 for monthly and 16 for weekly data. Retrieval-age limits are 1, 4, 35 and 8 days respectively for market, daily macro, monthly and weekly caches. Missing observations are never replaced with zero or estimates labeled as real data.

Price requests cover about three years. Failures retain prior data and its original retrieval time; old data is never relabeled as new. Market data consists of snapshots from different sources at different times, not a simultaneous live view of every global market.

6. Saving and reviewing

Saving records holdings, factor mappings, frozen locations, scenario shocks, trigger evidence, data dates, recent samples and price forecasts. Cloud records are isolated by account and show the latest 30. Delete records or compare up to three. Comparisons use each record’s original holdings, without attributing position changes to the scenario.

At maturity, forecasts are checked against the 21st or 30th valid daily observation after the origin. Incomplete intraday quotes are excluded. No score is produced before maturity or when data is missing. Evaluation covers representative asset prices only and does not validate an entire macro causal chain. This experimental model is not used for automated trading.

Anonymous usage statistics

We count visits, demo use and completed scenario sessions using a random browser identifier. We do not collect holdings, wallet addresses, email, full URLs or session recordings. Records expire after 90 days; returning visits mean a new session after 30 minutes of inactivity, not a verified person.

Applies to future events in this browser and domain. Privacy signals, local previews, Codex browsers and the signed-in administrator are automatically excluded.