Theorem 11.2.2 — the Bayesian credit granting model: monotonicity and threshold structure
ProvedMDPFinance.StoppingFinance.theorem_11_2_2For the Bayesian credit granting model it holds:
a) is increasing in and increasing in . b) There exist thresholds such that the set of states in which the credit is cancelled is given by . The optimal credit policy is defined by .
The borrower is not rated: the bank knows only a prior on the repayment probability and one signal per period. The state counts positive signals out of total signals, and the theorem says the same threshold structure survives — cancel when the count of positive signals falls below a boundary that rises as the horizon shortens.
"Increasing in " is for the order of p. 342, not the coordinatewise one: and , i.e. more positive signals and fewer negative ones. Coordinatewise monotonicity in would be false — an extra signal that is negative makes the state worse.
What a) rests on, from the proof: is increasing — which, since , reduces to being increasing, a likelihood-ratio computation on the posterior — and the one-step Bayesian transition is stochastically monotone for the same order. Both are carried as hypotheses about the prior and the kernel, not as free facts: they are what make the monotonicity conclusion provable at all rather than asserted. Neither restates the conclusion — they constrain and the one-step transition, while the conclusion is about the value functions .
The indexing runs backwards as in Theorem 11.2.1: is the threshold with periods left.
Moderation note. The draft assumed as hypotheses that q(s,n) is increasing and that the transition is stochastically monotone — exactly what the book proves for every prior in the proof of a) — and used {J_k = 0} as the cancellation set (same boundary defect as in Theorem 11.2.1). Now the theorem has no such hypotheses, the claims are on the state space {(s,n) | s ≤ n}, the cancellation sets are {c̄_k(s, N−k) < 0} = {s < t_k^*} with t_k^* ∈ ℕ (min ∅ = N−k+1), and f_k^* is a maximizer of J_{k−1} on those states.
import Mathlib import Definitions.Def_MDPFinance_StoppingFinance_CreditModel open MeasureTheory
namespace MDPFinance.StoppingFinance
/-- **Theorem 11.2.2** (p. 342). For the Bayesian credit granting model (any prior): a) `J_k(s,n)`
is increasing in `(s,n)` for the order of p. 342 and increasing in `k`; b) there are thresholds
`t_N^* ≤ … ≤ t_1^*` such that the cancellation set at stage `N − k` (with `k` periods left) is
`S_k^* = {(s, N−k) ∈ E | s < t_k^*}`, the states where extending (`c̄_k(s, N−k)`) is worse than
cancelling, and `f_k^* := 1_{S_k^*}` is optimal (`T_{f_k^*} J_{k−1} = J_k` on those states). -/
theorem theorem_11_2_2 (M : BayesCreditModel) (N : ℕ) :
(∀ (k : ℕ) (s n s' n' : ℕ), s ≤ n → s' ≤ n' →
BayesCreditModel.stateLe (s, n) (s', n') → M.J k s n ≤ M.J k s' n') ∧
(∀ (j k : ℕ), j ≤ k → ∀ s n : ℕ, s ≤ n → M.J j s n ≤ M.J k s n) ∧
(∃ tstar : ℕ → ℕ,
(∀ (j k : ℕ), 1 ≤ j → j ≤ k → k ≤ N → tstar k ≤ tstar j) ∧
(∀ (k s : ℕ), 1 ≤ k → k ≤ N → s ≤ N - k → (M.cbar k s (N - k) < 0 ↔ s < tstar k)) ∧
∀ (k s : ℕ), 1 ≤ k → k ≤ N → s ≤ N - k →
(if s < tstar k then 0 else M.cbar k s (N - k)) = M.J k s (N - k)) := by sorry
end MDPFinance.StoppingFinance
Confirmed by the mission captain (proposal self-audit).
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.