Quantile Bands vs Point Equity Price Targets
Why single-point stock targets hide risk, how pinball loss trains calibrated P10/P50/P90 bands across 1D-30D horizons, and why earnings widen tails.
Single-number stock price targets look tidy on a screen, yet they fail the moment a trader has to size a position or place a stop.
Say a model prints a 7-day forecast of $142.00 for a stock trading at $135.00. Without a spread around that number, two totally different market regimes look identical. In one regime, routine drift puts the 80% outcome envelope between $138.50 and $145.50. In another, an upcoming quarterly report stretches that same probability band from $118.00 to $164.00. A conditional mean forecast collapses both distributions into $142.00 and discards the downside geometry you actually need before putting capital on the line.
That gap is why Kuni, US-Market Forecaster plots three explicit quantile boundaries (P10, P50, P90) instead of a lone line.
The Single-Point Regression Trap
Standard regression models train against Mean Squared Error (MSE) or Root Mean Squared Error (RMSE). Squaring residuals pushes the model toward the conditional expected value:
MSE(y, y_hat) = (1 / N) * sum((y_i - y_hat_i)^2)
Equity returns break the cozy assumptions behind squared error in three ways:
- Skewed tails punish mean estimates. A single gap-down print pulls the arithmetic mean downward even when median session drift remains mildly positive.
- Symmetric penalties ignore risk asymmetry. Under MSE, missing
$5.00above the real close carries the same gradient penalty as missing$5.00below it. For a trader holding shares against a stop, downside errors matter far more than slightly undercounting a melt-up. - Volatility clusters. Volatility is never constant across weeks. A fixed error bar bolted onto a point prediction after training stays blind to whether implied turbulence just doubled ahead of a macro print or earnings call.
Pinball Loss Math for P10, P50, and P90
To estimate conditional percentiles directly rather than inferring Gaussian standard deviations around a mean, quantile models minimize pinball loss (also called check loss) for each target quantile q in (0, 1):
L_q(y, y_hat) = max(q * (y - y_hat), (q - 1) * (y - y_hat))
Written piecewise, the asymmetry becomes obvious:
If y >= y_hat: L_q(y, y_hat) = q * (y - y_hat)
If y < y_hat: L_q(y, y_hat) = (1 - q) * (y_hat - y)
Look at what happens at each of the three quantiles exposed on the Forecaster detail chart:
P50(q = 0.50): Both slopes equal0.5, reducing pinball loss to half the Mean Absolute Error (MAE). Half of historical realizations sit aboveP50and half sit below it. On stock cards, the percentage gap between spot price and the horizonP50endpoint powers the[Ai]median forecast badge.P90(q = 0.90): Under-predicting an upside move (y > y_hat) incurs a0.9xslope penalty, while over-predicting (y < y_hat) costs only0.1x. The optimizer settles where roughly 90% of realized closes remain at or below the upper dashed band.P10(q = 0.10): The slopes flip. Over-predicting a sell-off (y < y_hat) draws a heavy0.9xpenalty, pushing theP10support line down until roughly 90% of historical outcomes hold above it.
Monotonic Ordering and Conformal Calibration
Independent quantile heads can occasionally cross when noise spikes in thin regimes (for instance, raw P50 dipping briefly below P10 on a strange feature row). Enforcing non-crossing constraints—either structurally via cumulative positive offsets (P50 = P10 + softplus(delta_mid)) or post-hoc sorting—keeps the envelope physically coherent.
Even better, raw quantile tree or neural outputs still drift slightly miscalibrated out of sample during regime shifts. Split conformal quantile regression (CQR) checks empirical coverage residuals on a recent chronological calibration fold and adjusts the P10 and P90 offsets so the empirical [P10, P90] interval actually brackets ~80% of unseen trading days rather than drifting to 62% during stormy quarters.
Asymmetric Dispersion Across 1D to 30D Horizons
Under a naive textbook random walk, return dispersion widens strictly proportional to the square root of time (sigma * sqrt(t)). Real US equities rarely follow that neat rule across the five horizons in the Forecast Charts chapter:
| Horizon | Dominant Price Driver | Typical Corridor Shape | Operational Role |
|---|---|---|---|
1D |
Opening auction imbalance and overnight gaps | Tight symmetric envelope | Intraday entries and gap-fade bounds |
3D |
Short-swing catalyst drift | Slight skew toward active momentum | 2 to 3 session swing exits |
7D |
Weekly institutional flow and technical retests | Balanced baseline spread | Core swing setups |
14D |
Multi-week channel stability vs pullback risk | Wider lower tail if beta > 1.2 |
Position sizing and stop spacing |
30D |
Sector rotation and scheduled earnings prints | Broad asymmetric fan | Macro directional bias |
When a liquid mega-cap grinds upward on steady accumulation, P90 slopes steadily higher while P10 hugs rising moving-average support. Contrast that with a high-beta stock testing resistance below a flat 50-day EMA: P90 compresses near overhead supply while P10 drops sharply into the lower air pocket. Reading the ratio (P90 - Spot) / (Spot - P10) tells you immediately whether a trade offers asymmetric payoff or terrible risk-reward, long before you look at any composite score.
Why Earnings Windows Expand Tail Quantiles
Every automated forecast runs into a structural wall four times a year: quarterly earnings releases.
Over 60 quiet trading sessions, realized volatility might sit near 1.4% daily. Then a company reports after the close on day 61 and gaps 14% overnight on revenue guidance. No backward-looking 20-day moving average or RSI print sees an unannounced guidance cut coming.
Good quantile models respond to calendar proximity and pre-event implied dispersion by widening the outer P10 and P90 rails sharply into the reporting window while keeping P50 conservative. That behavior looks dramatic on a chart—suddenly your stop reference sits much lower and your stretch profit ceiling leaps higher—but the geometry is doing its job honestly. It signals that binary event risk has temporarily overwhelmed smooth technical drift.
Even with wide quantiles, automated bands have clear limits. They cannot anticipate FDA rulings, antitrust filings, emergency debt offerings, or overnight geopolitical shocks. Treat P10 as a structural invalidation threshold rather than a guaranteed floor, and inspect the Target Tracker guide alongside historical drawdown figures before trusting any single setup.