Barberis-Shleifer-Vishny Model

Barberis-Shleifer-Vishny Model

The Barberis, Shleifer & Vishny (1998) “Model of Investor Sentiment” is a single-representative-investor model that jointly generates short-run momentum, long-run reversal, post-earnings drift, and the value premium from one set of biased beliefs. Earnings actually follow a random walk, but the investor wrongly believes they are governed by one of two hidden regimes, switching slowly between them.

The setup is as follows: a risk-neutral representative investor prices a single asset whose true earnings process is the random walk
$N_{t+1}=N_t+y_{t+1},y_{t+1}=\{\begin{array}{ll}+y& \text{w.p. }1/2\\ -y& \text{w.p. }1/2\end{array}$
so the rational forecast is ${\mathbb{E}}_t[N_{t+1}]=N_t$. The investor instead believes earnings shocks $y_t$ are produced by one of two hidden Markov regimes that capture two distinct cognitive biases:

• Regime $1$ (mean-reverting): shocks are negatively autocorrelated, $\mathrm{Pr}(y_{t+1}=+y\mid y_t=+y)={\pi }_L$ with $0<{\pi }_L<1/2$. This regime captures conservatism bias — the investor under-weights single shocks because she expects them to revert.
• Regime $2$ (trending): shocks are positively autocorrelated, $\mathrm{Pr}(y_{t+1}=+y\mid y_t=+y)={\pi }_H$ with $1/2<{\pi }_H<1$. This regime captures the representativeness heuristic — the investor extrapolates from short streaks.

The model assumes that the investor’s prior puts more weight on Regime $1$ (i.e., the mean-reverting regime is “more typical”). Regime transitions follow a $2\times 2$ Markov chain with transition probabilities ${\lambda }_1$ (MR $\to$ T) and ${\lambda }_2$ (T $\to$ MR), with ${\lambda }_1,{\lambda }_2$ small and ${\lambda }_1<{\lambda }_2$. Each period the investor observes $y_t$ and Bayes-updates her posterior $q_t=\mathrm{Pr}(\text{regime MR}\mid \text{history})$.

The pricing equation that comes out is
$P_t=\frac{N_t}{\delta }+y_t(p_1-p_2{q}_t),$
where $\delta$ is the discount rate and $p_1,p_2$ are positive constants. The price equals the rational random-walk benchmark $N_t/\delta$ plus a sentiment term that swings with the investor’s posterior over regimes.

Mechanism in two phases. When earnings flatline at $100$ and then jump to $110$, the investor’s prior on Regime $1$ dominates, the new shock is partially “explained away” as mean-reverting noise, and her forecast for next period is below the rational $110$. The price under-reacts and post-earnings drift emerges. When earnings continue to rise to $120,130,140$, the streak shifts $q_t$ toward Regime $2$, the investor extrapolates the trend, and her forecast for next period now exceeds the rational $140$. The price over-reacts and a long-horizon reversal eventually corrects it. This is also true for the negative direction.

Predictions. $1$–$12$ month continuation, $13$–$60$ month reversal, post-earnings announcement drift, the value-glamour effect, and a sharper experimental prediction: investors should react more strongly to a price change preceded by fewer reversals (Bloomfield & Hales, 2001 confirms this; Poteshman, 2001 shows the analog in options markets reacting more strongly to volatility shocks that go in the same direction several days in a row). Within the over- and under-reaction family, BSV is the single representative investor alternative to the heterogeneous-agent Hong-Stein model.

Intuition

The puzzle BSV solves is sharp. The same cross-section of stocks displays positive return autocorrelation at one-year horizons (continuation) and negative return autocorrelation at three-to-five-year horizons (reversal). A single mechanism that produces only continuation cannot also produce reversal, and vice versa. BSV’s answer is that the investor isn’t wrong about prices in any one direction — she is confused about which earnings process is generating the data, and her confusion resolves slowly enough that her error flips sign over time.

The two regimes are deliberately picked to mirror two famous cognitive biases. Conservatism (Edwards, 1968) is the tendency to update beliefs too slowly when given new information; the mean-reverting regime captures this because under it, a single positive shock should be partially undone next period, so the investor’s forecast lags the rational benchmark. The representativeness heuristic (Kahneman & Tversky) is the tendency to see patterns in short sequences and infer that they will continue; the trending regime captures this because under it, the investor takes a streak as evidence that the underlying process has positive autocorrelation. The investor’s prior leans toward conservatism — she enters with $\mathrm{Pr}(\text{MR})>1/2$ — which is why short news streaks produce under-reaction. Long news streaks eventually shift the posterior toward the trending regime, and at that point her forecasts overshoot.

The model is elegant because it gets both halves of the empirical pattern from a single free parameter trajectory — the path of $q_t$ — driven entirely by the realized history of earnings. The investor never changes her model of the world; she just updates which of two regimes she thinks is currently in force. With only the mean-reverting regime, the model would predict pure under-reaction and continuation forever. With only the trending regime, it would predict pure over-reaction and immediate overshoot. The interaction of both regimes, weighted by the investor’s gradually shifting posterior, is what generates continuation followed by reversal.

Properties

(Posterior Drives Prices) Prices deviate from the rational benchmark by a sentiment term proportional to the current shock times the current posterior on the mean-reverting regime.

The pricing equation $P_t=N_t/\delta +y_t(p_1-p_2{q}_t)$ has two parts. The first is the rational random-walk price, which moves one-for-one with current earnings. The second is the sentiment correction. When $q_t$ is high (investor believes the world is mean-reverting), the bracket $p_1-p_2{q}_t$ is small or negative, so the price under-reacts to $y_t$ — the investor expects the shock to undo itself. When $q_t$ is low (investor believes the world is trending), the bracket is large, the price over-reacts, and the investor expects the shock to continue. The full empirical content of the model — momentum, post-earnings drift, reversal, value premium — falls out of how $q_t$ evolves in response to the realized earnings path.

(Streak Length Matters More Than Magnitude) The shift in $q_t$ depends on how consistent the recent earnings sign has been, not on how large any single shock was.

This is the model’s most distinctive testable prediction. A single $+y$ shock barely moves $q_t$ — the investor explains it away as Regime-1 noise. A streak of three or four consecutive $+y$ shocks moves $q_t$ substantially because that pattern is much more likely under Regime $2$ than under Regime $1$. The empirical signature is that markets should react to streak structure in a way pure random-walk pricing would not. Bloomfield & Hales (2001) confirm this in an experimental asset market: MBA students shown a random-walk earnings process price stocks more aggressively when an earnings change is preceded by fewer reversals, exactly as BSV predicts. Poteshman (2001) finds an analog in options markets reacting more strongly to volatility shocks when several go in the same direction in a row.

(Why ${\lambda }_1<{\lambda }_2$ Is Load-Bearing) The mean-reverting regime is set up to be more persistent than the trending regime, ensuring under-reaction is the default state.

The transition probability ${\lambda }_1$ governs the mean-reverting → trending switch and ${\lambda }_2$ the reverse. Setting ${\lambda }_1<{\lambda }_2$ encodes the assumption that the investor finds it harder to escape the mean-reverting regime than to fall back into it; in equilibrium her posterior $q_t$ stays above $1/2$ most of the time. Without this asymmetry, the model would predict overshoot and reversal but not the empirically dominant pattern of initial under-reaction. The asymmetry is what makes momentum the primary phenomenon and reversal the secondary correction, matching the observed magnitudes — momentum signals are stronger and more reliable than reversal signals in the data.

(Continuation and Reversal Are Joint, Not Separate) The same parameters ${\pi }_L,{\pi }_H,{\lambda }_1,{\lambda }_2$ generate both empirical patterns; they cannot be tuned separately.

This is the central contribution. Earlier behavioral stories explained continuation or reversal but had to invoke unrelated mechanisms for the other. BSV ties them together: over a typical earnings path, the investor’s posterior starts high (mean-reverting), drifts down through a streak, triggers overshoot near the end of the streak, and then snaps back when reversion arrives. The post-earnings drift, the $1$–$12$ month momentum window, the $13$–$60$ month reversal window, and the value-glamour effect all emerge from this single trajectory. A complete behavioral account of momentum cannot just match the headline numbers — it must, like BSV, generate the joint pattern from one mechanism with shared parameters.

(Single-Agent Limits) BSV is silent on trading volume, on cross-sectional variation tied to information environments, and on cross-asset spillovers.

Because there is only one investor, no one trades against anyone, so the model has no implications for trading volume — yet empirically, momentum and trading volume are positively correlated (Lee & Swaminathan, 2000). Because the investor sees only public earnings, BSV cannot explain why momentum is stronger in low-analyst-coverage stocks (Hong, Lim & Stein, 2000). Because there is only one asset and one investor, BSV says nothing about why momentum spills across markets and asset classes the way time-series momentum documents. These silences are the gap that the Hong-Stein model is designed to fill.

Examples

Posterior trajectory: how a streak of $+$ shocks moves $q_t$

Take the parameter values used in the original BSV illustration: ${\pi }_L=1/3$, ${\pi }_H=3/4$, ${\lambda }_1=1/10$, ${\lambda }_2=3/10$, with $q_0=0.5$. The investor watches a $20$-period earnings sign sequence and Bayes-updates $q_t$ each period.

$t$	$y_t$	$q_t$		$t$	$y_t$	$q_t$
$0$	$+$	$0.50$		$11$	$+$	$0.74$
$1$	$-$	$0.80$		$12$	$+$	$0.56$
$2$	$+$	$0.90$		$13$	$+$	$0.44$
$3$	$-$	$0.93$		$14$	$+$	$0.36$
$4$	$+$	$0.94$		$15$	$-$	$0.74$
$5$	$+$	$0.74$		$16$	$+$	$0.89$
$6$	$-$	$0.89$		$17$	$+$	$0.69$
$7$	$-$	$0.69$		$18$	$-$	$0.87$
$8$	$+$	$0.87$		$19$	$+$	$0.92$
$9$	$-$	$0.92$		$20$	$+$	$0.72$

Two patterns matter. First, single-period reversals push $q_t$ up sharply — a flip from $+$ to $-$ (or vice versa) is exactly what the mean-reverting regime predicts, so the investor’s belief in MR strengthens. Second, streaks of consecutive $+$ shocks (periods $11$–$14$) drag $q_t$ down monotonically toward $0.36$, because the streak is increasingly hard to explain under MR. By period $14$ the investor has moved from “definitely MR” to “probably trending,” and her forecast for period $15$ is now extrapolative. The single $-$ at period $15$ then snaps her belief back to $0.74$ — one observation flips her mental regime entirely. The price implications track this trajectory: under-reaction during periods $0$–$4$, gradual incorporation through periods $5$–$11$, overshoot during periods $11$–$14$, and the snap-back reversal at period $15$.

The Bloomfield-Hales experimental test

Bloomfield & Hales (2001) ran an experimental asset market with MBA students. Subjects were shown an earnings process that the experimenters constructed to be a true random walk and asked to set prices on a fictional stock period by period. The key manipulation was the number of recent reversals in the displayed earnings history before the current change.

Under rational pricing, the number of recent reversals should be irrelevant — earnings follow a random walk, the next shock is independent of the past, and the price should move one-for-one with the current shock. Under BSV, however, the number of recent reversals controls $q_t$: a history with many reversals leaves the investor confident she is in Regime $1$ (high $q_t$) and therefore under-reactive to the next shock; a history with few reversals shifts her toward Regime $2$ (low $q_t$) and makes her over-reactive.

The data match BSV. Subjects priced changes preceded by fewer reversals more aggressively than changes preceded by many reversals, exactly as the regime-switching machinery predicts. This is a clean test because the random-walk rational benchmark makes the opposite prediction (no effect of streak structure), so any positive effect of streak length is direct evidence for BSV-style regime confusion. Poteshman (2001) found the same pattern in options markets reacting to volatility shocks.

Why BSV gets post-earnings drift "for free"

Post-earnings announcement drift is the empirical fact that stocks with positive earnings surprises continue to outperform for $30$–$60$ trading days after the announcement, and stocks with negative surprises continue to underperform. This is one of the most robust anomalies in finance and is hard to rationalize as risk premium because the drift direction is known once the announcement is made.

BSV produces post-earnings drift mechanically. When a single earnings surprise arrives, the investor’s $q_t$ is high (she has a strong prior on the mean-reverting regime), so she partially “explains away” the surprise as noise that will revert. Her current-period price moves only fractionally toward the rational level. As subsequent periods bring no reversal — quiet news, slight drift in the same direction — her $q_t$ slowly drifts down, the under-reaction unwinds, and the price drifts toward fundamental value. The drift rate is the rate at which $q_t$ updates, which the model parameters ${\pi }_L,{\pi }_H,{\lambda }_1,{\lambda }_2$ pin down.

The same mechanism that produces $1$–$12$ month momentum produces post-earnings drift on a faster timescale. This is what makes BSV parsimonious: a single behavioral parameter set generates both anomalies, with no extra assumptions, just by varying the time horizon over which $q_t$ updates.

See Also

• 8.9. Reaction Models — parent hub: BSV vs Hong-Stein comparison.
• 8.9.2. Hong-Stein Model — sibling: heterogeneous-agent alternative.
• 8.8. Momentum — the empirical phenomenon BSV explains.
• 8.8.1. Cross-Sectional Momentum — the headline anomaly target.
• 8.6.1. Conservatism — the cognitive bias underlying Regime $1$.
• 8.6.5. Representativeness Heuristic — the cognitive bias underlying Regime $2$.
• De Bondt & Thaler 1985 — the long-horizon reversal that BSV jointly explains.
• 8.3. Challenges to Rational Pricing — efficient-markets framework BSV departs from.