FPUS1 model validation
Does FPUS1's predicted risk match the risk that actually showed up? Each day, before the open, the model forecasts the volatility of many portfolios. The bias statistic compares those forecasts with what happened: 1.0 means the forecasts were right on average, below 1 means the model was too cautious (it predicted more risk than occurred), and above 1 means it underpredicted risk.
Point-in-time forecasts began Jul 31, 2026 (66 sessions so far); those are the only out-of-sample results. Earlier forecasts were computed retroactively, and the model's settings were chosen on that same 2020–2026 history, so the backfilled period is in-sample.
Summary
| Portfolios | Count | Live bias | Live in band | Backfilled bias | Backfilled in band |
|---|---|---|---|---|---|
| Random portfolios 40 books drawn at random each month from the 1,500 largest stocks: half long-only, half long/short. | 40 | 0.99 | 98% | 0.98 | 88% |
| Hedge-fund portfolios The 50 largest hedge-fund portfolios reported in 13F filings (funds are not named). | 50 | 0.91 | 96% | 0.98 | 80% |
| Cap-weighted market Every stock in the estimation universe, weighted by market cap. | 1 | 0.91 | 100% | 0.98 | 80% |
| Equal-weighted market Every stock in the estimation universe, equally weighted. | 1 | 1.01 | 100% | 0.97 | 87% |
| Factor portfolios The model's own long/short factor portfolios. | 24 | 0.88 | 71% | 0.89 | 73% |
| Sector portfolios Cap-weighted portfolios of each broad industry division. | 8 | 1.15 | 75% | 1.08 | 64% |
Live: one bias value per portfolio over every point-in-time session since Jul 31, 2026; the median across portfolios, and the share of portfolios inside 0.8–1.2. With only 66 sessions, a calibrated model will still land anywhere in roughly that range by chance. Backfilled: rolling 63-day windows over 2020–Jul 2026; the median, and the share of windows inside 0.8–1.2.
Why sector portfolios run above 1: apart from its six themes, FPUS1 has no industry factors, so most of the risk of a concentrated sector book is treated as stock-specific risk that should diversify away. For a book of one industry it doesn't, so the model underpredicts its risk by about 10%, both live and in the backfilled history. Adding industry factors is a planned improvement for the next model.
Bias over time
Random portfolios
Hedge-fund portfolios
Cap-weighted market
Equal-weighted market
Factor portfolios
Sector portfolios
Hardest periods
The 63-day windows furthest from 1.0, all in the backfilled period.
| Window ending | Portfolio | Bias | |
|---|---|---|---|
| Apr 11, 2025 | Sector portfolio (resources) | 1.88 | risk underpredicted |
| Apr 10, 2025 | Sector portfolio (construction) | 1.76 | risk underpredicted |
| Apr 10, 2025 | A random portfolio | 1.67 | risk underpredicted |
| Apr 10, 2025 | A large hedge-fund portfolio | 1.62 | risk underpredicted |
| Nov 12, 2020 | Semiconductors factor | 1.60 | risk underpredicted |
| Sep 11, 2020 | Sector portfolio (transport util) | 0.41 | risk overpredicted |
| Sep 11, 2020 | A random portfolio | 0.43 | risk overpredicted |
| Apr 10, 2025 | Cap-weighted market | 1.55 | risk underpredicted |
Coverage and stability
- 2,030 stocks forecast on Nov 02, 2026; stock-specific risk for 3.4% of market value (newly listed names) uses a fallback estimate.
How stable each factor's exposures are month to month
Median correlation of each stock's exposure with its exposure a month earlier. Values near 1 mean slow-moving characteristics; lower values mean the factor turns over faster.
| Factor | Month-to-month correlation |
|---|---|
| International | 1.00 |
| Size | 0.99 |
| Hedge-Fund Ownership* | 0.99 |
| Leverage | 0.96 |
| Dividend Yield | 0.96 |
| Growth | 0.94 |
| Beta | 0.93 |
| Profitability | 0.92 |
| Long-Term Momentum | 0.90 |
| Liquidity | 0.88 |
| Morning Activity | 0.87 |
| Value | 0.86 |
| Short Interest | 0.85 |
| Oil | 0.74 |
| Residual Volatility | 0.72 |
| Gold | 0.68 |
| China | 0.65 |
| Treasury (Duration) | 0.65 |
| Bitcoin / Crypto | 0.62 |
| Semiconductors | 0.62 |
| Short-Sale Activity | 0.16 |
| Medium-Term Momentum† | 0.02 |
| One-Day Momentum† | 0.01 |
| Short-Term Momentum† | 0.00 |
* The underlying 13F holdings data changes only four times a year, unless funds file in between, so most months compare nearly identical inputs and the median overstates how stable the factor is from quarter to quarter.
† The momentum factors measure returns over a month or less (medium-term momentum looks back about 20 trading days), so a month later the window has fully rolled over and a correlation near zero is expected. It is how these factors are built, not instability in the model.