| # Environment Simulator Logic |
|
|
| This document defines the complete simulation framework for the CFO environment, including **backend state tracking**, **financial statement generation**, **fundraising success simulation**, and **stochastic evaluation design**. All modules share a common monthly time axis `t = 0, 1, 2, ..., T`, and the agent only observes information up to the current time `t`. |
|
|
| --- |
|
|
| ## Module A: Backend Tracking (Ledger) |
|
|
| ### Purpose |
|
|
| Maintain a running ledger of the company's financial state at each month. The system tracks **two parallel views**: |
|
|
| 1. **P&L Ledger** (accrual basis) — revenue and costs recognized when *earned/incurred* |
| 2. **Cash Ledger** (cash basis) — actual money entering/leaving the bank account |
|
|
| The ledger is **append-only** and the agent can only read `[0, t]` — no future information is visible. |
|
|
| **Simulation boundaries** (from `config.json`): |
| - Start date: `company_config.initial_date` (e.g., 2005-01-01) |
| - Max duration: `environment_config.max_episode_months` (e.g., 300 months) |
| - The simulation runs from `t = 0` to at most `t = max_episode_months`, or until cash goes negative. |
|
|
| ### A.1 Core State Variables |
|
|
| At each month `t`, the backend tracks: |
|
|
| **Business State:** |
|
|
| | Field | Description | Source / Logic | |
| |-------|-------------|----------------| |
| | `t` | Month index (0-based) | System clock | |
| | `date` | Calendar month-end date | `combined_economic_data_2015_2025.csv` | |
| | `active_users` | Monthly active borrowers | Simulated (see A.2) | |
| | `net_new_users` | New borrowers this month | `Users(t) − Users(t-1)` | |
| | `loan_portfolio_gross` | Total outstanding loans (gross) | Running balance (see A.6) | |
| | `allowance` | Allowance for credit losses | Running balance (see A.6) | |
| | `lending_rate` | Annual lending rate | `Tsy2Y(t)/100 + Baa_Yield(t)/100` | |
|
|
| **P&L (Accrual Basis):** |
|
|
| | Field | Description | Source / Logic | |
| |-------|-------------|----------------| |
| | `revenue` | Interest income accrued | `Loan_Portfolio_Gross × Lending_Rate / 12` | |
| | `credit_loss_provision` | Expected loan losses | `Revenue × (1 − Gross_Margin) × Provision_Share` | |
| | `cogs` | Cost of goods sold | `Revenue × (1 − Gross_Margin)` | |
| | `gross_profit` | Gross profit | `Revenue − COGS` | |
| | `opex` | Operating expenses | `Gross_Profit − EBITDA` | |
| | `ebitda` | EBITDA | `Revenue × EBITDA_Margin` | |
| | `interest_expense` | Interest on company's debt | Sum of `principal × rate / 12` per debt instrument | |
| | `net_income` | Bottom line | `EBITDA − Interest_Expense − Taxes` | |
|
|
| **Cash (Bank Account):** |
|
|
| | Field | Description | Source / Logic | |
| |-------|-------------|----------------| |
| | `cash_balance` | Actual bank balance | Previous + all cash inflows − all cash outflows | |
| | `cash_in_interest` | Borrower interest payments received | `Revenue(t-1) × Collection_Rate` (lagged) | |
| | `cash_in_principal` | Borrower principal repayments | `Loan_Portfolio_Gross(t-1) / Avg_Loan_Term × Collection_Rate` (lagged) | |
| | `cash_out_originations` | New loans funded | `max(0, Net_New_Users(t)) × Avg_Loan_Size` (immediate) | |
| | `cash_out_servicing` | Servicing cost payments | `Servicing_Costs(t-1)` (lagged) | |
| | `cash_out_opex` | Operating cost payments | `OpEx(t-1)` (lagged) | |
| | `cash_out_debt_interest` | Debt interest paid | Same month | |
| | `cash_out_debt_repayment` | Debt principal paid | Scheduled payment | |
| | `cash_in_fundraising` | Funds received from raises | Same month (if success) | |
|
|
| **Balance Sheet Running Balances:** |
|
|
| | Field | Description | Source / Logic | |
| |-------|-------------|----------------| |
| | `interest_receivable` | Accrued revenue awaiting collection | `Revenue(t)` — cleared next month | |
| | `principal_receivable` | Scheduled repayment awaiting cash | `Loan_Portfolio_Gross(t) / Avg_Loan_Term` | |
| | `accounts_payable` | Accrued costs awaiting payment | `OpEx(t) + Servicing_Costs(t)` — paid next month | |
| | `write_offs` | Bad loans removed from books | `Credit_Loss_Provision(t-4)` — non-cash | |
|
|
| **Ownership & Capital:** |
|
|
| | Field | Description | Source / Logic | |
| |-------|-------------|----------------| |
| | `total_debt` | Outstanding debt principal | Cumulative from fundraising events | |
| | `total_equity_raised` | Cumulative equity raised | Cumulative from fundraising events | |
| | `shares_outstanding` | Current total shares | Updated on equity raises | |
|
|
| ### A.2 Revenue & User Simulation |
|
|
| Revenue and users evolve monthly based on indicators from `combined_economic_data_2015_2025.csv`. |
|
|
| > **Adjusted Columns Convention:** The simulation uses `adj_Monthly_User_Growth` (not the original `Monthly_User_Growth`) to drive user growth. The `adj_` columns contain amplified growth bumps at three checkpoint periods (months 23-36, 53-62, 106-117) that create cash liquidity traps for the lending company. Original columns are preserved in the CSV for reference. Similarly, fundraising probabilities use `adj_P_debt` and `adj_P_equity` from `fundraising_success_probabilities_2015_2025.csv`. |
|
|
| **Active Users (Borrowers):** |
|
|
| ``` |
| Users(t) = Users(t-1) × (1 + adj_Monthly_User_Growth(t) / 100) |
| Net_New(t) = Users(t) − Users(t-1) |
| ``` |
|
|
| **Loan Portfolio (Gross) — Running Balance:** |
|
|
| The gross loan portfolio is tracked as a running balance (see A.6 for full logic): |
|
|
| ``` |
| Loan_Portfolio_Gross(t) = Loan_Portfolio_Gross(t-1) |
| + New_Originations(t) |
| − Scheduled_Principal_Repayment(t) |
| − Write_Offs(t) |
| ``` |
|
|
| > `Avg_Loan_Size` = $10,000 (fixed, from `config.json → company_config.average_loan_size`) |
| > At t=0: `Loan_Portfolio_Gross(0) = initial_customers × Avg_Loan_Size` |
|
|
| **Revenue (Interest Income) — Accrual:** |
|
|
| The company earns interest on its gross loan portfolio. The lending rate = risk-free base rate + credit spread: |
|
|
| ``` |
| Lending_Rate(t) = Tsy2Y(t) / 100 + Baa_Yield(t) / 100 |
| Revenue(t) = Loan_Portfolio_Gross(t) × Lending_Rate(t) / 12 |
| ``` |
|
|
| > `Tsy2Y(t)` = 2-Year Treasury yield from `combined_economic_data_2015_2025.csv` |
| > `Baa_Yield(t)` = ICE BofA Corporate Bond OAS (credit spread) from `combined_economic_data_2015_2025.csv` |
| > `loan_term_years` = 2 (from `config.json → company_config.loan_term_years`) |
| > Divide by 12 because rates are annual but revenue is monthly |
| > |
| > **Why Tsy2Y?** The base rate matches the loan duration — 2-year loans use the 2-Year Treasury yield as the risk-free benchmark. |
| > |
| > **Why Baa_Yield?** This is the market credit spread (OAS), not an absolute yield. It captures the risk premium the market demands for lending — widens during crises (e.g., 2.58% in Mar 2020), tightens in calm periods (e.g., 0.77% in Sep 2025). Both components are fully dynamic from the environment. |
| |
| **Cost Breakdown — Accrual:** |
| |
| ``` |
| COGS(t) = Revenue(t) × (1 − Gross_Margin(t) / 100) |
| └─ Credit_Loss_Provision(t) = COGS(t) × Provision_Share |
| └─ Servicing_Costs(t) = COGS(t) × (1 − Provision_Share) |
| Gross_Profit(t) = Revenue(t) − COGS(t) |
| EBITDA(t) = Revenue(t) × EBITDA_Margin(t) / 100 |
| OpEx(t) = Gross_Profit(t) − EBITDA(t) |
| ``` |
| |
| > `Provision_Share` = 0.40 (from `config.json`) — 40% of COGS is credit loss provision (non-cash), 60% is servicing costs |
| |
| ### A.3 Accrual vs. Cash: Timing Lags |
| |
| A lending company has significant timing differences between when items appear on the P&L and when cash actually moves. This section explains each lag. |
| |
| #### Interest Revenue → Cash Collection (lag: **1 month**) |
| |
| Interest income is *accrued* on the P&L in the month it's earned. But cash arrives when borrowers make their monthly payment, which involves processing time (ACH settlement, payment posting). Additionally, not all borrowers pay — some are delinquent or default. |
| |
| ``` |
| Cash_In_Interest(t) = Revenue(t-1) × Collection_Rate |
| ``` |
| |
| > `Collection_Rate` = 0.96 (96%, from `config.json`) — 4% of accrued interest is never collected |
| > At `t=0`, there is no prior month, so `Cash_In_Interest(0) = 0` |
| |
| #### Principal Repayments → Cash In (lag: **1 month**) |
| |
| When borrowers make monthly payments, part goes to interest (above) and part repays principal. Principal repayments are NOT revenue — they reduce the loan asset on the balance sheet. But they are real cash inflows. |
| |
| ``` |
| Monthly_Principal_Rate = 1 / Avg_Loan_Term_Months |
| Cash_In_Principal(t) = Loan_Portfolio_Gross(t-1) × Monthly_Principal_Rate × Collection_Rate |
| ``` |
| |
| > `Avg_Loan_Term_Months` = 24 (from `config.json`) — 2-year loan term |
| > This is cash coming back from the existing portfolio each month |
| |
| #### New Loan Originations → Cash Out (lag: **0 months, immediate**) |
| |
| When new borrowers are onboarded, the company must fund their loans immediately from its bank account. This is the biggest cash drain for a growing lender — it is NOT an expense on the P&L (it becomes an asset), but it consumes cash. |
| |
| ``` |
| Cash_Out_Originations(t) = max(0, Net_New_Users(t)) × Avg_Loan_Size |
| ``` |
| |
| > If users decline (negative growth), no origination cash outflow occurs |
| > This creates the key tension: **fast growth = high P&L revenue but heavy cash burn** |
|
|
| #### Credit Loss Provision → Write-off (lag: **4 months**, non-cash) |
|
|
| Credit losses are *provisioned* on the P&L when loans are originated or risk is assessed. This is a **non-cash** accrual that increases the Allowance for Credit Losses (a contra-asset on the Balance Sheet). |
|
|
| When a borrower becomes 90-120+ days delinquent (roughly 4 months later), the loan is **written off**. A write-off is a **Balance Sheet reclassification only** — it does NOT cause cash to leave the bank: |
|
|
| ``` |
| Write_Off(t) = Credit_Loss_Provision(t-4) |
| |
| Balance Sheet entry: |
| Loan Portfolio (Gross) ↓ by Write_Off amount |
| Allowance for Credit Losses ↓ by Write_Off amount |
| Loan Portfolio (Net) unchanged (gross and allowance both decrease) |
| ``` |
|
|
| > At early months (`t < 4`), no write-offs have matured yet, so `Write_Off = 0` |
| > The cash impact of defaults is already captured through `Collection_Rate < 100%`: borrowers who default stop making interest and principal payments, so `Cash_In_Interest` and `Cash_In_Principal` are reduced accordingly. |
|
|
| #### Operating Expenses → Cash Out (lag: **1 month**) |
|
|
| OpEx includes salaries, technology, marketing, rent, etc. Salaries are paid same-month, but vendor invoices are typically net-30. Blended across all OpEx: |
|
|
| ``` |
| Cash_Out_OpEx(t) = OpEx(t-1) |
| ``` |
|
|
| > At `t=0`, `Cash_Out_OpEx(0) = 0` |
|
|
| #### Servicing Costs → Cash Out (lag: **1 month**) |
|
|
| The non-provision portion of COGS (payment processing, customer support, loan servicing): |
|
|
| ``` |
| Cash_Out_Servicing(t) = Servicing_Costs(t-1) |
| ``` |
|
|
| > `servicing_lag_months` = 1 (from `config.json`) |
|
|
| #### Company Debt Interest → Cash Out (lag: **0 months, same month**) |
|
|
| Interest the company owes on its own debt (from fundraising) is accrued and paid in the same month. |
|
|
| ``` |
| Cash_Out_Debt_Interest(t) = Interest_Expense(t) |
| ``` |
|
|
| #### Company Debt Repayment → Cash Out (lag: **0 months, scheduled**) |
|
|
| Principal repayments on the company's debt follow the amortization schedule. |
|
|
| ``` |
| Cash_Out_Debt_Repayment(t) = scheduled amount per debt_instruments table |
| ``` |
|
|
| > Repayment type: `environment_config.debt_repayment_type` (default: `"amortizing"`) |
| > Default maturity: `environment_config.debt_maturity_months` (default: 36 months) |
|
|
| #### Fundraising Proceeds → Cash In (lag: **0 months, immediate**) |
|
|
| When a fundraising round succeeds, cash arrives immediately. |
|
|
| ``` |
| Cash_In_Fundraising(t) = Amount_Raised (if success, else 0) |
| ``` |
|
|
| #### Summary Table |
|
|
| | P&L Item (Accrual) | Cash Event | Lag | Rate/Adjustment | |
| |---------------------|------------|-----|-----------------| |
| | Revenue (interest accrued) | Borrower interest payments received | **t+1** | × Collection_Rate (96%) | |
| | *(not on P&L)* | Borrower principal repayments received | **t+1** | × Collection_Rate (96%) | |
| | *(not on P&L)* | New loan originations funded | **t+0** | immediate, full amount | |
| | Credit Loss Provision | Write-off (BS reclassification, **non-cash**) | **t+4** | Gross ↓, Allowance ↓, Net unchanged | |
| | Servicing Costs (COGS) | Servicing payments | **t+1** | accrued amount | |
| | OpEx | Vendor/salary payments | **t+1** | accrued amount | |
| | Interest Expense (debt) | Debt interest paid | **t+0** | same month | |
| | *(not on P&L)* | Debt principal repayment | **t+0** | per schedule | |
| | *(not on P&L)* | Fundraising proceeds | **t+0** | if success | |
|
|
| ### A.4 Cash Balance Update |
|
|
| Each month, the **actual bank balance** is updated using **cash-basis** items only: |
|
|
| ``` |
| Cash(t) = Cash(t-1) |
| |
| + Cash_In_Interest(t) Revenue(t-1) × Collection_Rate |
| + Cash_In_Principal(t) Loan_Portfolio_Gross(t-1) / Avg_Loan_Term × Collection_Rate |
| + Cash_In_Fundraising(t) if fundraising succeeds this month |
| |
| − Cash_Out_Originations(t) Net_New_Users(t) × Avg_Loan_Size |
| − Cash_Out_Servicing(t) Servicing_Costs(t-1) |
| − Cash_Out_OpEx(t) OpEx(t-1) |
| − Cash_Out_Debt_Interest(t) Interest_Expense(t) |
| − Cash_Out_Debt_Repayment(t) per amortization schedule |
| ``` |
|
|
| > **Key insight:** A lending company can be **profitable on the P&L but cash-negative** if it grows quickly. New loan originations drain cash immediately, while interest income trickles in over months. This forces the CFO agent to balance growth ambition with liquidity management — and fundraise at the right time. |
|
|
| ### A.5 Storage Schema |
|
|
| **P&L Ledger** (accrual basis, one row per month-end): |
|
|
| ``` |
| pnl_ledger ( |
| t INT PRIMARY KEY, |
| date DATE, -- month-end date (e.g., 2015-01-31) |
| active_users BIGINT, |
| net_new_users BIGINT, |
| lending_rate DECIMAL, |
| |
| -- P&L items (accrual) |
| revenue DECIMAL, |
| credit_loss_provision DECIMAL, |
| servicing_costs DECIMAL, |
| cogs DECIMAL, |
| gross_profit DECIMAL, |
| opex DECIMAL, |
| ebitda DECIMAL, |
| interest_expense DECIMAL, |
| net_income DECIMAL, |
| |
| -- Balance Sheet running balances |
| loan_portfolio_gross DECIMAL, -- running balance (see A.6) |
| allowance DECIMAL, -- running balance (see A.6) |
| interest_receivable DECIMAL, -- Revenue(t), cleared at t+1 |
| principal_receivable DECIMAL, -- scheduled repayment, cleared at t+1 |
| accounts_payable DECIMAL, -- OpEx(t) + Servicing_Costs(t), cleared at t+1 |
| write_offs DECIMAL -- Credit_Loss_Provision(t-4), non-cash |
| ) |
| ``` |
|
|
| **Cash Ledger** (cash basis, one row per month-end): |
|
|
| ``` |
| cash_ledger ( |
| t INT PRIMARY KEY, |
| date DATE, -- month-end date (e.g., 2015-01-31) |
| cash_in_interest DECIMAL, -- Revenue(t-1) × Collection_Rate |
| cash_in_principal DECIMAL, -- portfolio principal repayments |
| cash_in_fundraising DECIMAL, -- 0 unless fundraising succeeds |
| cash_out_originations DECIMAL, -- new loans funded |
| cash_out_servicing DECIMAL, -- Servicing_Costs(t-1) |
| cash_out_opex DECIMAL, -- OpEx(t-1) |
| cash_out_debt_interest DECIMAL, -- same month |
| cash_out_debt_repayment DECIMAL, -- per schedule |
| net_cash_flow DECIMAL, -- sum of all above |
| cash_balance DECIMAL -- running bank balance |
| ) |
| ``` |
|
|
| **Auxiliary tables:** |
|
|
| ``` |
| debt_instruments ( |
| id INT PRIMARY KEY, |
| issued_at INT, -- month t when issued |
| principal DECIMAL, |
| rate DECIMAL, -- annual interest rate |
| maturity_months INT, |
| remaining_principal DECIMAL |
| ) |
| |
| equity_rounds ( |
| id INT PRIMARY KEY, |
| issued_at INT, -- month t when issued |
| amount_raised DECIMAL, |
| shares_issued BIGINT, |
| price_per_share DECIMAL, |
| investor_label TEXT |
| ) |
| |
| capital_summary ( |
| t INT PRIMARY KEY, |
| total_debt DECIMAL, |
| total_equity_raised DECIMAL, |
| shares_outstanding BIGINT |
| ) |
| ``` |
|
|
| ### A.6 Loan Portfolio & Allowance Tracking |
|
|
| The loan portfolio and allowance are tracked as **running balances** to ensure the Balance Sheet always balances. |
|
|
| **Loan Portfolio (Gross):** |
|
|
| ``` |
| Loan_Portfolio_Gross(t) = Loan_Portfolio_Gross(t-1) |
| + New_Originations(t) |
| − Scheduled_Principal_Repayment(t) |
| − Write_Offs(t) |
| ``` |
|
|
| Where: |
| - `New_Originations(t)` = `max(0, Net_New_Users(t)) × Avg_Loan_Size` |
| - `Scheduled_Principal_Repayment(t)` = `Loan_Portfolio_Gross(t-1) / Avg_Loan_Term_Months` |
| - `Write_Offs(t)` = `Credit_Loss_Provision(t - 4)` (bad loans removed from books after ~4 months delinquent) |
|
|
| **Allowance for Credit Losses:** |
|
|
| ``` |
| Allowance(t) = Allowance(t-1) |
| + Credit_Loss_Provision(t) provision added (from P&L) |
| − Write_Offs(t) used up when loan is written off |
| ``` |
|
|
| > When a write-off occurs, **both** Gross and Allowance decrease by the same amount, so **Net is unchanged**. This is a reclassification, not a new loss — the loss was already recognized when the provision was booked. |
|
|
| **Interest Receivable:** |
|
|
| ``` |
| Interest_Receivable(t) = Revenue(t) |
| ``` |
|
|
| > Accrued this month, collected next month. Cleared when Cash_In_Interest arrives at t+1. |
| > Uncollected portion (1 − Collection_Rate) is absorbed by the Allowance. |
| |
| **Principal Receivable:** |
| |
| ``` |
| Principal_Receivable(t) = Scheduled_Principal_Repayment(t) |
| ``` |
| |
| > Scheduled this month, cash arrives next month. Uncollected portion absorbed by Allowance. |
| |
| **Accounts Payable:** |
| |
| ``` |
| Accounts_Payable(t) = OpEx(t) + Servicing_Costs(t) |
| ``` |
| |
| > Accrued this month, paid next month (1-month lag). |
| |
| ### A.7 Initial Balance Sheet (t = 0) |
| |
| At simulation start, the company begins with a pre-existing state. All values must satisfy `A = L + E`: |
| |
| ``` |
| Assets: |
| Cash = initial_cash (from config.json) |
| Interest Receivable = 0 (no prior accrual) |
| Principal Receivable = 0 (no prior schedule) |
| Loan Portfolio (Gross) = initial_customers × Avg_Loan_Size |
| Allowance = 0 (no provisions yet) |
| Loan Portfolio (Net) = Loan Portfolio (Gross) |
| ──────────────────────────────── |
| Total Assets = initial_cash + initial_customers × Avg_Loan_Size |
|
|
| Liabilities: |
| Total Debt = initial_debt (from config.json, default 0) |
| Accounts Payable = 0 (no prior accrual) |
| ──────────────────────────────── |
| Total Liabilities = initial_debt |
|
|
| Equity: |
| Paid-in Capital = Total Assets − Total Liabilities (founders funded everything) |
| Retained Earnings = 0 (no prior income) |
| ──────────────────────────────── |
| Total Equity = Paid-in Capital |
|
|
| Cap Table: |
| Shares Outstanding = initial_shares_outstanding (from config.json, default 10,500,000) |
| Implied Share Price = initial_share_price (from config.json, default $10.0) |
|
|
| Balance check: A = L + E ✓ |
| ``` |
| |
| > With `initial_cash = $15M` and `initial_customers = 5,000 × $10K = $50M`, Total Assets = $65M, Paid-in Capital = $65M. |
| > `initial_shares_outstanding = 10,000,000` — matches pre-ESOP Cap Table (Founders 8M + Seed 2M). Implied share price $6.50 × 10M = $65M = Paid-in Capital. |
| |
| ### A.8 Balance Sheet Reconciliation |
| |
| Every monthly event must maintain `A = L + E`. Here is how each event type preserves the equation: |
| |
| **1. New loan origination** (Cash → Loan Portfolio): |
| ``` |
| Cash ↓ $X | Loan Portfolio (Gross) ↑ $X |
| Assets net: unchanged | L+E: unchanged ✓ |
| ``` |
| |
| **2. Interest revenue accrual** (P&L → Balance Sheet): |
| ``` |
| Interest Receivable ↑ $Y (Asset ↑) |
| Revenue → Net Income → RE ↑ $Y (Equity ↑) |
| A ↑ = E ↑ ✓ |
| ``` |
| |
| **3. Interest cash collection at t+1** (Receivable → Cash): |
| ``` |
| Cash ↑ $Y × 96% (Asset ↑) |
| Interest Receivable ↓ $Y (Asset ↓) |
| Uncollected $Y × 4%: |
| Allowance ↓ $Y × 4% (contra-asset ↓ = Asset ↑) |
| ... net: absorbed by existing allowance |
| A net: unchanged | L+E: unchanged ✓ |
| ``` |
| |
| **4. Credit loss provision** (P&L → Allowance): |
| ``` |
| Allowance ↑ $Z (contra-asset ↑ = Net Assets ↓) |
| COGS → Net Income → RE ↓ $Z (Equity ↓) |
| A ↓ = E ↓ ✓ |
| ``` |
| |
| **5. Loan write-off** (BS reclassification, non-cash): |
| ``` |
| Loan Portfolio (Gross) ↓ $W (Asset ↓) |
| Allowance ↓ $W (contra-asset ↓ = Asset ↑) |
| Loan Portfolio (Net): unchanged (↓ and ↑ cancel) |
| A: unchanged | L+E: unchanged ✓ |
| ``` |
| |
| **6. OpEx / Servicing accrual** (P&L → AP): |
| ``` |
| Accounts Payable ↑ $V (Liability ↑) |
| OpEx → Net Income → RE ↓ $V (Equity ↓) |
| L ↑ by $V, E ↓ by $V → L+E: unchanged | A: unchanged ✓ |
| ``` |
| |
| **7. OpEx / Servicing cash payment at t+1**: |
| ``` |
| Cash ↓ $V (Asset ↓) |
| Accounts Payable ↓ $V (Liability ↓) |
| A ↓ = L ↓ ✓ |
| ``` |
| |
| **8. Principal repayment scheduled** (Loan → Receivable): |
| ``` |
| Loan Portfolio (Gross) ↓ $P (Asset ↓) |
| Principal Receivable ↑ $P (Asset ↑) |
| A: unchanged | L+E: unchanged ✓ |
| ``` |
| |
| **9. Principal cash collection at t+1**: |
| ``` |
| Cash ↑ $P × 96% (Asset ↑) |
| Principal Receivable ↓ $P (Asset ↓) |
| Uncollected $P × 4%: |
| Allowance ↓ $P × 4% (absorbed by existing allowance) |
| A net: unchanged | L+E: unchanged ✓ |
| ``` |
| |
| **10. Fundraising (equity):** |
| ``` |
| Cash ↑ $F (Asset ↑) |
| Paid-in Capital ↑ $F (Equity ↑) |
| A ↑ = E ↑ ✓ |
| ``` |
| |
| **11. Fundraising (debt):** |
| ``` |
| Cash ↑ $F (Asset ↑) |
| Total Debt ↑ $F (Liability ↑) |
| A ↑ = L ↑ ✓ |
| ``` |
| |
| **12. Debt interest payment** (same month, accrual = cash): |
| ``` |
| Cash ↓ $I (Asset ↓) |
| Interest Expense → RE ↓ $I (Equity ↓) |
| A ↓ = E ↓ ✓ |
| ``` |
| |
| **13. Debt principal repayment:** |
| ``` |
| Cash ↓ $D (Asset ↓) |
| Total Debt ↓ $D (Liability ↓) |
| A ↓ = L ↓ ✓ |
| ``` |
| |
| ### A.9 Visibility Rule |
| |
| At time `t`, the agent can query: |
| - `pnl_ledger WHERE t <= current_t` |
| - `cash_ledger WHERE t <= current_t` |
| - `capital_summary WHERE t <= current_t` |
| - `debt_instruments WHERE issued_at <= current_t` |
| - `equity_rounds WHERE issued_at <= current_t` |
| |
| No future rows are ever exposed. |
| |
| --- |
| |
| ## Module B: Financial Statement Simulation |
| |
| ### Purpose |
| |
| Generate the three core financial statements (**Income Statement**, **Balance Sheet**, **Cash Flow Statement**) on demand, triggered by specific agent actions. |
| |
| ### B.1 Triggers |
| |
| There are two actions that trigger financial statement generation: |
| |
| | Action | Triggers | Output | |
| |--------|----------|--------| |
| | `book_closing` | End-of-period close | 3 financial statements (IS, BS, CF) | |
| | `fund_raising_request` | Fundraising attempt | 3 financial statements + updated Cap Table | |
| |
| ### B.2 Action: `book_closing` |
| |
| **When:** Agent decides to close the books at current month `t`. |
| |
| **Reporting period:** Always **Year-to-Date (YTD)** — from January of the current year through the current month. For example, if `t` falls in August 2017, the reporting period is Jan 2017 – Aug 2017. |
| |
| ``` |
| ytd_start = January of the year that date(t) falls in |
| ytd_end = date(t) |
| ``` |
| |
| **Process:** |
| |
| ``` |
| 1. Determine YTD period: |
| ytd_start = first month of the current calendar year |
| ytd_end = current month t |
|
|
| 2. Freeze both ledgers at month t |
|
|
| 3. Generate: |
| a. Income Statement (YTD: ytd_start to t) |
| b. Balance Sheet (snapshot: as of t) |
| c. Cash Flow Statement (YTD: ytd_start to t) |
|
|
| 4. Store statements as a versioned snapshot |
|
|
| 5. Return statements to the agent |
| ``` |
| |
| **Income Statement** (YTD: `[ytd_start, t]`): |
| |
| | Line Item | Source | |
| |-----------|--------| |
| | Revenue | `SUM(pnl_ledger.revenue)` from ytd_start to t | |
| | Cost of Goods Sold | `SUM(pnl_ledger.cogs)` from ytd_start to t | |
| | Gross Profit | Revenue − COGS | |
| | Operating Expenses | `SUM(pnl_ledger.opex)` from ytd_start to t | |
| | EBITDA | Gross Profit − OpEx | |
| | Interest Expense | `SUM(pnl_ledger.interest_expense)` from ytd_start to t | |
| | Pre-tax Income | EBITDA − Interest Expense | |
| | Tax (simplified %) | Pre-tax Income × tax_rate | |
| | Net Income | Pre-tax Income − Tax | |
| |
| > Example: book_closing in March 2018 → Income Statement covers Jan–Mar 2018 |
| > Example: book_closing in December 2018 → full-year Income Statement for 2018 |
| |
| **Balance Sheet** (snapshot at `t`): |
| |
| | Line Item | Source | |
| |-----------|--------| |
| | **Assets** | | |
| | Cash & Equivalents | `cash_ledger.cash_balance` at t | |
| | Interest Receivable | `Revenue(t)` — accrued this month, cash arrives at t+1 | |
| | Principal Receivable | `Loan_Portfolio_Gross(t) / Avg_Loan_Term` — scheduled repayment, cash arrives at t+1 | |
| | Loan Portfolio (Gross) | Running balance (see A.6) | |
| | Less: Allowance for Credit Losses | Running balance (see A.6) | |
| | Loan Portfolio (Net) | Gross − Allowance | |
| | **Total Assets** | Sum of above | |
| | **Liabilities** | | |
| | Total Debt | `SUM(debt_instruments.remaining_principal)` at t | |
| | Accounts Payable | `OpEx(t) + Servicing_Costs(t)` — accrued this month, paid at t+1 | |
| | **Total Liabilities** | Sum of above | |
| | **Equity** | | |
| | Paid-in Capital | Initial capital + `SUM(equity_rounds.amount_raised)` | |
| | Retained Earnings | Cumulative `SUM(pnl_ledger.net_income)` from t=0 to t | |
| | **Total Equity** | Paid-in Capital + Retained Earnings | |
| |
| > Balance Sheet is always a point-in-time snapshot — not affected by the YTD period. |
| > Balance check: Total Assets = Total Liabilities + Total Equity |
| > |
| > **Why Interest Receivable?** Revenue is accrued in month t (P&L), but cash arrives in t+1. Without this asset, the balance sheet would have equity (via retained earnings) increasing without a matching asset increase. Interest Receivable bridges the 1-month gap. |
| > |
| > **Why Principal Receivable?** Same logic — principal repayments are scheduled at t but cash arrives at t+1. This is the portion of the loan portfolio that is "due this month" and waiting for payment processing. |
| |
| **Cash Flow Statement** (YTD: `[ytd_start, t]`, built from `cash_ledger`): |
| |
| | Section | Line Items | Source | |
| |---------|-----------|--------| |
| | **Operating Activities** | | | |
| | Interest received from borrowers | `SUM(cash_in_interest)` ytd_start to t | cash_ledger | |
| | Servicing & OpEx paid | `−SUM(cash_out_servicing + cash_out_opex)` ytd_start to t | cash_ledger | |
| | *Net Cash from Operations* | *Sum of above* | | |
| | **Investing Activities** | | | |
| | New loans originated | `−SUM(cash_out_originations)` ytd_start to t | cash_ledger | |
| | Principal repayments received | `+SUM(cash_in_principal)` ytd_start to t | cash_ledger | |
| | *Net Cash from Investing* | *Sum of above* | | |
| | **Financing Activities** | | | |
| | Debt raised | `+SUM(cash_in_fundraising)` where type=debt, ytd_start to t | cash_ledger | |
| | Equity raised | `+SUM(cash_in_fundraising)` where type=equity, ytd_start to t | cash_ledger | |
| | Debt interest paid | `−SUM(cash_out_debt_interest)` ytd_start to t | cash_ledger | |
| | Debt principal repaid | `−SUM(cash_out_debt_repayment)` ytd_start to t | cash_ledger | |
| | *Net Cash from Financing* | *Sum of above* | | |
| | **Net Change in Cash** | Operating + Investing + Financing | | |
| | **Beginning Cash** | `cash_ledger.cash_balance` at ytd_start − 1 (prior year-end) | | |
| | **Ending Cash** | Beginning + Net Change | | |
| |
| > **Cross-check with Balance Sheet:** Ending Cash on the CF Statement MUST equal Cash & Equivalents on the Balance Sheet. Both are sourced from `cash_ledger.cash_balance` at t. |
| > |
| > **Cross-check with Income Statement:** Net Income (IS) + non-cash items (provision) − working capital changes (ΔReceivables, ΔPayables) ≈ Net Cash from Operations (CF). This is the indirect method relationship and can be verified as a sanity check. |
| |
| ### B.3 Action: `fund_raising_request` |
| |
| **When:** Agent requests debt or equity fundraising. |
| |
| **Input parameters from agent:** |
| |
| | Parameter | Description | |
| |-----------|-------------| |
| | `type` | `"debt"` or `"equity"` | |
| | `amount` | Requested amount | |
| | `terms` | Rate / valuation / maturity (depending on type) | |
| |
| **Process:** |
| |
| ``` |
| 1. Generate YTD financial statements (same as book_closing, steps 1-5) |
| → These represent the company's state "as presented to investors" |
| |
| 2. Determine fundraising success: |
| a. Look up P_debt or P_equity for current month t |
| (from Module C: Fundraising Success Simulation) |
| b. Draw random number r ~ Uniform(0, 1) |
| c. success = (r < P_success) |
|
|
| 3. If SUCCESS: |
| a. For DEBT: |
| - Insert into debt_instruments (principal, rate, maturity) |
| - Update ledger.total_debt |
| - Update ledger.cash_balance += amount_raised |
| b. For EQUITY: |
| - Calculate new shares: shares_new = amount / price_per_share |
| - Insert into equity_rounds |
| - Update shares_outstanding += shares_new |
| - Update cap_table (recalculate all ownership %) |
| - Update ledger.total_equity_raised |
| - Update ledger.cash_balance += amount_raised |
| |
| 4. If FAILURE: |
| - No changes to ledger, debt_instruments, or equity_rounds |
| - Return failure notice to agent |
| - Agent may retry in a future month |
| |
| 5. Return: |
| - Financial statements (always) |
| - Fundraising result (success/failure) |
| - Updated cap table (if equity success) |
| ``` |
| |
| ### B.4 Cap Table Update Logic (Equity Only) |
| |
| On a successful equity raise: |
| |
| ``` |
| new_shares = amount_raised / price_per_share |
| total_shares_after = shares_outstanding + new_shares |
| |
| For each existing shareholder: |
| new_ownership% = their_shares / total_shares_after |
| |
| New investor: |
| ownership% = new_shares / total_shares_after |
| ``` |
| |
| **Example:** |
| |
| | Before | Shares | % | | After ($2M at $20M val) | Shares | % | |
| |--------|--------|---|-|------------------------|--------|---| |
| | Founders | 8,000,000 | 76% | | Founders | 8,000,000 | 72.7% | |
| | Seed | 2,000,000 | 19% | | Seed | 2,000,000 | 18.2% | |
| | ESOP | 500,000 | 5% | | ESOP | 500,000 | 4.5% | |
| | | | | | Series A | 500,000 | 4.5% | |
| | **Total** | **10,500,000** | **100%** | | **Total** | **11,000,000** | **100%** | |
| |
| ### B.5 Statement Versioning |
| |
| Each generation is stored with metadata: |
| |
| ``` |
| financial_snapshots ( |
| snapshot_id INT PRIMARY KEY, |
| trigger TEXT, -- 'book_closing' or 'fund_raising_request' |
| t INT, -- month index when generated |
| date DATE, -- month-end date (e.g., 2018-03-31) |
| ytd_start_date DATE, -- YTD period start (e.g., 2018-01-31) |
| fiscal_year INT, -- e.g., 2018 |
| months_in_period INT, -- e.g., 3 for Jan-Mar |
| income_stmt JSON/BLOB, -- YTD: ytd_start to t |
| balance_sheet JSON/BLOB, -- Point-in-time snapshot at t |
| cashflow_stmt JSON/BLOB, -- YTD: ytd_start to t |
| cap_table JSON/BLOB, -- NULL for book_closing |
| balance_check BOOLEAN -- Total Assets = Total Liabilities + Total Equity |
| ) |
| ``` |
| |
| > Example: book_closing called in March 2018 → |
| > `date = 2018-03-31`, `ytd_start_date = 2018-01-31`, `fiscal_year = 2018`, `months_in_period = 3` |
|
|
| --- |
|
|
| ## Module C: Fundraising Success Simulation |
|
|
| Fundraising probability is determined by two layers: |
|
|
| 1. **Macro layer** — base probability from CSV, driven by market conditions (VIX, interest rates) |
| 2. **Company-state layer** — modifiers that reduce probability based on the company's own financial state |
|
|
| ``` |
| P_adjusted = P_base_csv × company_modifier |
| ``` |
|
|
| ### C.1 Macro Layer: Base Probabilities |
|
|
| #### Debt |
| - Key driver: **SOFR / Fed Funds** (lower rate → higher success probability) |
| - Use SOFR when available (Apr 2018+), fallback to FEDFUNDS |
|
|
| ``` |
| P_debt_base = 1 − (Rate − Rate_min) / (Rate_max − Rate_min) |
| ``` |
|
|
| #### Equity |
| - Key driver: **VIX** (lower VIX → higher success probability) |
|
|
| ``` |
| P_equity_base = 1 − (VIX − VIX_min) / (VIX_max − VIX_min) |
| ``` |
|
|
| ### C.2 Company-State Layer: Difficulty Progression |
|
|
| #### Equity: Round-Count Decay |
|
|
| Each completed equity round makes the next one harder (models "Series A is easier than Series D"): |
|
|
| ``` |
| equity_modifier = equity_round_decay ^ num_completed_equity_rounds |
| P_equity = P_equity_base × equity_modifier |
| ``` |
|
|
| With `equity_round_decay = 0.75` (configurable in `environment_config`): |
|
|
| | Completed Rounds | Modifier | Effect | |
| |------------------|----------|--------| |
| | 0 (first raise) | 1.00x | No penalty | |
| | 1 | 0.75x | | |
| | 2 | 0.56x | | |
| | 3 | 0.42x | | |
| | 4 | 0.32x | | |
| | 5 | 0.24x | | |
| | 6+ | ~0.18x | Nearly impossible | |
|
|
| #### Debt: Leverage Penalty on Probability |
|
|
| High leverage ratio (total_debt / total_equity) reduces debt approval odds. Below a safe threshold, no penalty applies; above it, the modifier declines linearly to zero: |
|
|
| ``` |
| leverage_ratio = total_debt / (paid_in_capital + retained_earnings) |
| |
| if leverage_ratio <= debt_safe_leverage: |
| debt_modifier = 1.0 |
| else: |
| excess = leverage_ratio - debt_safe_leverage |
| debt_modifier = max(0, 1 - debt_leverage_prob_decay × excess) |
| |
| P_debt = P_debt_base × debt_modifier |
| ``` |
|
|
| With `debt_safe_leverage = 0.5` and `debt_leverage_prob_decay = 1.5`: |
|
|
| | Leverage | Modifier | Effect | |
| |----------|----------|--------| |
| | 0.0–0.5 | 1.00x | Safe zone, no penalty | |
| | 0.7 | 0.70x | | |
| | 1.0 | 0.25x | | |
| | 1.17+ | 0.00 | Impossible | |
|
|
| #### Debt: Leverage Spread on Cost |
|
|
| Higher leverage increases the interest rate the company pays, on top of the market rate: |
|
|
| ``` |
| base_rate = (Tsy2Y + Baa_Yield) / 100 |
| leverage_spread = max(0, leverage_ratio - debt_safe_leverage) × debt_leverage_spread_bps / 10000 |
| cost_rate = base_rate + leverage_spread |
| ``` |
|
|
| With `debt_leverage_spread_bps = 500` (5% per unit of excess leverage): |
|
|
| | Leverage | Spread | Effect | |
| |----------|--------|--------| |
| | 0.3 | +0.0% | Below safe threshold | |
| | 0.7 | +1.0% | | |
| | 1.0 | +2.5% | | |
|
|
| ### C.3 Simulate Success |
|
|
| For each fundraising attempt at month `t`: |
|
|
| ```python |
| prob = P_base_csv × company_modifier # adjusted probability |
| roll = random() |
| success = roll < prob |
| ``` |
|
|
| ### C.4 Amount Raised & Cost |
|
|
| **Amount Raised (on success):** |
|
|
| ``` |
| fill_rate = Uniform(0.7, 1.0) |
| Amount_Raised = Fund_Ask × fill_rate |
| ``` |
|
|
| **Cost:** |
|
|
| ``` |
| Debt_Cost = (Tsy2Y + Baa_Yield) / 100 + leverage_spread (annual interest rate, includes company risk) |
| Equity_Cost = Fund_Ask / Implied_Valuation (dilution %) |
| ``` |
|
|
| ### C.5 Deferred Settlement |
|
|
| Fundraising results are not immediate. Upon submission, the outcome (success/failure) and delivery month are determined but not revealed to the agent: |
|
|
| ``` |
| delivery_month = current_month + randint(fundraising_delivery_min, fundraising_delivery_max) |
| ``` |
|
|
| The agent receives a "submitted" confirmation. At the delivery month, a notification is sent with the result: |
| - **If approved:** funds are deposited, balance sheet updated, cap table refreshed (for equity) |
| - **If declined:** no changes, agent may retry |
|
|
| Config parameters: `fundraising_delivery_min = 1`, `fundraising_delivery_max = 6` (in `environment_config`). |
|
|
| ### C.6 Configuration Parameters |
|
|
| All fundraising difficulty parameters are in `config.json → environment_config`: |
|
|
| ```json |
| "equity_round_decay": 0.75, |
| "debt_safe_leverage": 0.5, |
| "debt_leverage_prob_decay": 1.5, |
| "debt_leverage_spread_bps": 500, |
| "fundraising_delivery_min": 1, |
| "fundraising_delivery_max": 6 |
| ``` |
|
|
| ### C.7 Historical Reference (2015–2025) |
|
|
| #### Normalization Parameters |
|
|
| | Metric | Min | Max | |
| |--------|-----|-----| |
| | Rate (SOFR/FEDFUNDS) | 0.01% | 5.34% | |
| | VIX | 10.13 | 57.74 | |
|
|
| #### Summary Statistics |
|
|
| | Metric | P_debt | P_equity | |
| |--------|--------|----------| |
| | Min | 0.00 | 0.00 | |
| | Max | 1.00 | 1.00 | |
| | Mean | 0.63 | 0.83 | |
|
|
| #### Notable Periods |
|
|
| | Event | Date | Probability | Driver | |
| |-------|------|-------------|--------| |
| | **Best for Debt** | Apr 2021 | P_debt = 1.00 | Rate = 0.01% (near-zero rates) | |
| | **Worst for Debt** | Jul 2024 | P_debt = 0.00 | Rate = 5.34% (peak rates) | |
| | **Best for Equity** | Oct 2017 | P_equity = 1.00 | VIX = 10.13 (lowest volatility) | |
| | **Worst for Equity** | Mar 2020 | P_equity = 0.00 | VIX = 57.74 (COVID crash) | |
|
|
| > Data source: `fundraising_success_probabilities_2015_2025.csv` |
|
|
| --- |
|
|
| ## Module D: Stochastic Simulation & Evaluation Design |
|
|
| ### D.1 Deterministic vs. Stochastic Components |
|
|
| 当前模拟器中各环节可分为两类: |
|
|
| #### 确定性逻辑(Deterministic)— 不应加入随机性 |
|
|
| 这些要素由环境数据或会计公式直接驱动,是"游戏规则"本身,加入随机性会破坏逻辑一致性: |
|
|
| | 组件 | 原因 | |
| |------|------| |
| | 宏观经济指标 (Tsy2Y, Baa_Yield, VIX, FEDFUNDS 等) | 来自真实历史数据,是所有 agent 面对的相同环境 | |
| | Lending Rate = Tsy2Y + Baa_Yield | 纯公式,由环境决定 | |
| | Revenue = Portfolio × Rate / 12 | 会计恒等式,给定 portfolio 和 rate 后无歧义 | |
| | COGS / Gross Profit / EBITDA / OpEx 分解 | 给定 margin 后是纯算术 | |
| | 所有 Timing Lag 结构 (1 month, 4 month) | 系统规则,不应随机化 | |
| | Cash Balance 更新公式 | 会计恒等式 | |
| | 三表生成与配平 (IS, BS, CF) | 会计规则,不容差异 | |
| | P_debt / P_equity 计算 | 由环境指标确定的概率值本身是确定的 | |
|
|
| #### 可引入随机性(Stochastic)— 增加 episode 间差异 |
|
|
| 这些要素在现实中本身存在波动,加入噪声可以测试 agent 在不确定环境中的鲁棒性: |
|
|
| | 组件 | 当前设计 | 随机化方案 | 理由 | |
| |------|----------|-----------|------| |
| | **Fundraising 成功/失败** | `random() < P_success` | 保持不变(已经是随机的) | 核心随机性来源 | |
| | **Collection Rate** | 固定 0.96 | `clip(N(0.96, σ₁), 0.85, 1.0)` 每月独立抽样 | 真实中催收率受经济环境和借款人行为波动 | |
| | **User Growth** | 直接用 CSV 中的 `Monthly_User_Growth` | `Monthly_User_Growth(t) + N(0, σ₂)` | 用户增长有行业噪声,同一经济环境下不同公司有差异 | |
| | **Gross Margin** | 直接用 CSV 中的 `Gross_Margin` | `Gross_Margin(t) + N(0, σ₃)` | 信用损失和服务成本有月度波动 | |
| | **EBITDA Margin** | 直接用 CSV 中的 `EBITDA_Margin` | `EBITDA_Margin(t) + N(0, σ₄)` | 运营效率有随机波动 | |
| | **Fundraising 到账金额** | `Fund_Ask × P_success` (全额或零) | `Fund_Ask × Uniform(0.7, 1.0)` (部分到账) | 现实中很少恰好拿到全额 | |
|
|
| > 注意:Lending Rate、Revenue 公式、Cash Formula、三表逻辑**不加噪声**。这些是确定性的"物理定律"。随机性只加在输入参数上,公式本身保持精确。 |
|
|
| ### D.2 噪声参数 (Noise Parameters) |
|
|
| 所有噪声参数集中管理,写入 `config.json → stochastic_config`: |
|
|
| ``` |
| stochastic_config: { |
| "enabled": true, -- 开关:false 时退化为完全确定性 |
| "seed": null, -- 全局种子,null 表示使用 episode_id |
| "n_episodes": 1000, -- 每个 agent 的运行次数 |
| |
| "collection_rate_std": 0.04, -- σ₁: Collection Rate 标准差 |
| "user_growth_std": 0.5, -- σ₂: Monthly User Growth 标准差 (百分点) |
| "gross_margin_std": 2.0, -- σ₃: Gross Margin 标准差 (百分点) |
| "ebitda_margin_std": 1.5, -- σ₄: EBITDA Margin 标准差 (百分点) |
| "fundraising_fill_range": [0.7, 1.0] -- 到账比例的均匀分布范围 |
| } |
| ``` |
|
|
| ### D.3 可比性保证:Common Random Numbers (CRN) |
|
|
| **核心问题:** 加入随机性后,如何确保 Agent A 和 Agent B 的结果是可比的? |
|
|
| **方案:种子控制 + 配对比较 (Paired Comparison with CRN)** |
|
|
| ``` |
| 对于 episode i = 1, 2, ..., 1000: |
| seed_i = base_seed + i |
| rng_i = RandomGenerator(seed_i) |
| |
| 用 rng_i 预生成该 episode 的所有随机序列: |
| collection_rates[0..T] ← N(0.96, σ₁) per month |
| user_growth_noise[0..T] ← N(0, σ₂) per month |
| gross_margin_noise[0..T] ← N(0, σ₃) per month |
| ebitda_margin_noise[0..T] ← N(0, σ₄) per month |
| fundraising_rolls[0..T] ← Uniform(0, 1) per month (用于判定成功/失败) |
| fundraising_fills[0..T] ← Uniform(0.7, 1.0) per month |
| |
| Agent A 和 Agent B 在 episode i 中面对**完全相同**的随机序列 |
| 唯一的差异来自 agent 的决策(何时融资、融多少、选 debt 还是 equity) |
| ``` |
|
|
| > **关键:** 随机序列是环境的属性,不是 agent 的属性。同一 episode 中所有 agent 面对相同的"天气",区别只在于他们如何应对。 |
|
|
| **为什么这有效?** |
|
|
| 1. **消除运气差异:** 如果 Agent A 碰巧遇到高 Collection Rate 的 episode 而 Agent B 没有,比较不公平。CRN 确保两者面对同样的随机环境 |
| 2. **降低所需样本量:** 配对比较的方差远小于独立比较。Var(A-B|paired) << Var(A) + Var(B)。1000 次可能就足够,而独立比较可能需要 10000+ |
| 3. **可复现:** 给定 seed,任何 episode 都可以精确复现,便于 debug |
|
|
| ### D.4 评估指标与统计方法 |
|
|
| **每个 episode 产出一个 score:** |
|
|
| ``` |
| Score_i = (TTM_Revenue × Valuation_Multiple) + Cash_Balance if cash never went negative |
| Score_i = 0 if cash went negative at any point |
| ``` |
|
|
| > `Valuation_Multiple` = 10 (from `config.json → environment_config.valuation_multiple`) |
| > TTM = Trailing Twelve Months revenue (sum of last 12 months from `pnl_ledger.revenue`) |
|
|
| **跨 episode 聚合:** |
|
|
| | 指标 | 计算 | 用途 | |
| |------|------|------| |
| | Mean Score | `mean(Score_1..Score_N)` | 主要排名指标 | |
| | Survival Rate | `count(Score > 0) / N` | agent 避免现金危机的能力 | |
| | Std Dev | `std(Score_1..Score_N)` | 策略稳定性 | |
| | Median Score | `median(Score_1..Score_N)` | 抗极端值的稳健指标 | |
| | 5th Percentile | `quantile(Score, 0.05)` | 最差情况表现(tail risk) | |
|
|
| **Agent 间比较:** |
|
|
| ``` |
| 对于每个 episode i: |
| Δ_i = Score_A_i − Score_B_i (配对差值) |
| |
| 统计检验: |
| H₀: mean(Δ) = 0 (两个 agent 无差异) |
| H₁: mean(Δ) ≠ 0 |
| |
| 使用 paired t-test 或 Wilcoxon signed-rank test |
| 报告 p-value 和 95% confidence interval for mean(Δ) |
| ``` |
|
|
| > 因为使用了 CRN,配对差值 Δ_i 的方差很小,即使两个 agent 的平均表现差异不大也能检测出来。 |
| |
| ### D.5 实验流程 |
| |
| ``` |
| 输入: |
| agents = [Agent_A, Agent_B, Agent_C, ...] |
| N = 1000 |
| base_seed = 42 |
| |
| 流程: |
| for i in 1..N: |
| env_i = generate_environment(seed = base_seed + i) |
| → 预生成所有月份的随机参数 |
| → 加载确定性的宏观数据 |
| |
| for agent in agents: |
| reset agent state |
| score_i[agent] = run_episode(agent, env_i) |
| |
| 输出: |
| for each agent: |
| mean, std, median, survival_rate, p5, p95 |
| |
| for each agent pair (A, B): |
| paired_diff = score[A] − score[B] |
| t_stat, p_value = paired_ttest(paired_diff) |
| → "Agent A is significantly better/worse than Agent B (p < 0.05)" |
| ``` |
| |
| --- |
|
|
| ## System Flow Diagram |
|
|
| ``` |
| ──── Every month (automatic) ──── |
| │ |
| ▼ |
| Module A: update both ledgers |
| ┌──────────────────────────────────┐ |
| │ 1. Compute Users(t), Revenue(t) │ ← P&L ledger (accrual) |
| │ COGS(t), OpEx(t), etc. │ |
| │ │ |
| │ 2. Compute cash flows with lags │ ← Cash ledger (bank account) |
| │ Revenue(t-1) → cash in │ |
| │ New loans(t) → cash out │ |
| │ OpEx(t-1) → cash out │ |
| │ │ |
| │ 3. Update BS running balances │ ← Loan Portfolio, Allowance, |
| │ Provision(t-4) → write-off │ Receivables, Payables |
| │ (non-cash BS reclassification)│ |
| │ │ |
| │ 4. Update cash_balance │ |
| └──────────────────────────────────┘ |
| │ |
| ▼ |
| Agent decides action at month t |
| │ |
| ├──── action = "book_closing" |
| │ │ |
| │ ▼ |
| │ Freeze both ledgers [0, t] |
| │ │ |
| │ ▼ |
| │ Generate 3 statements |
| │ (IS from pnl_ledger, |
| │ BS from both, |
| │ CF from cash_ledger) |
| │ │ |
| │ ▼ |
| │ Return to agent |
| │ |
| └──── action = "fund_raising_request" |
| │ |
| ▼ |
| Generate 3 statements (same as above) |
| │ |
| ▼ |
| Module C: evaluate P_success |
| │ |
| ┌────┴────┐ |
| ▼ ▼ |
| SUCCESS FAILURE |
| │ │ |
| ▼ │ |
| Update: │ |
| - cash_ledger │ |
| - capital_summary│ |
| - debt/equity │ |
| - cap table │ |
| │ │ |
| ▼ ▼ |
| Return results to agent |
| ``` |
|
|
| --- |
|
|
| ## Data Dependencies |
|
|
| | Module | Reads From | Writes To | |
| |--------|-----------|----------| |
| | **A: Tracking** | `data_paths.combined_econ_data` (Tsy2Y, Baa_Yield, Gross_Margin, EBITDA_Margin, **adj_Monthly_User_Growth**) + `company_config` (loan params, lag params) | `pnl_ledger`, `cash_ledger`, `capital_summary`, `debt_instruments`, `equity_rounds` | |
| | **B: Statements** | `pnl_ledger`, `cash_ledger`, `capital_summary`, `debt_instruments`, `equity_rounds` | `financial_snapshots` | |
| | **C: Fundraising** | `data_paths.fundraising_probs` (**adj_P_debt**, **adj_P_equity**) | Results fed into Module A & B | |
| | **D: Stochastic** | `stochastic_config` (noise params, seed, n_episodes) | Pre-generated random sequences per episode | |
| |
| --- |
| |
| ## Appendix: Accounting Review — Issues Found & Fixed |
| |
| 在设计三表配平检查时,发现并修正了以下三个问题: |
| |
| ### Issue 1: Write-off 被错误归类为现金流出 |
| |
| **原始设计:** `Cash_Out_Writeoffs(t) = Credit_Loss_Provision(t-4)` 作为现金流出项出现在 cash_ledger 和现金流量表中。 |
|
|
| **问题:** Write-off 是资产负债表内部的重新分类(Gross ↓, Allowance ↓, Net 不变),不会导致现金离开银行账户。违约对现金的影响已经通过 `Collection_Rate < 100%` 体现——违约借款人停止支付利息和本金,导致 `Cash_In_Interest` 和 `Cash_In_Principal` 减少。将 write-off 同时计入现金流出会造成**重复计算**。 |
|
|
| **修正:** 从 cash_ledger、A.4 现金公式、现金流量表中移除 `Cash_Out_Writeoffs`。Write-off 仅作为资产负债表非现金操作保留在 A.6 的 running balance 逻辑中。 |
| |
| ### Issue 2: 缺少 Interest Receivable 和 Principal Receivable |
| |
| **原始设计:** 资产负债表资产端只有 Cash、Loan Portfolio、Other Assets。 |
| |
| **问题:** Revenue 在月 t 通过 P&L 确认(accrual),增加 Retained Earnings(Equity ↑),但现金要到 t+1 才到账。如果资产端没有对应的 Receivable,则 Equity 增加了但 Asset 没有同步增加,**A ≠ L + E**。同理,本金还款在 t 月安排但 t+1 到账,Loan Portfolio Gross 已经减少但 Cash 尚未增加,中间需要 Principal Receivable 来桥接。 |
| |
| **修正:** 在资产负债表中新增 `Interest Receivable = Revenue(t)` 和 `Principal Receivable = Loan_Portfolio_Gross(t) / Avg_Loan_Term`,并在 A.6 中定义其清算逻辑。 |
| |
| ### Issue 3: Accounts Payable 不完整 |
| |
| **原始设计:** `Accounts Payable = OpEx(t)` — 仅包含运营费用。 |
| |
| **问题:** Servicing Costs(COGS 中的非 provision 部分)同样有 1 个月的现金滞后(月 t 确认,t+1 支付)。如果只计入 OpEx 的应付而遗漏 Servicing Costs,则 P&L 确认的 Servicing_Costs 减少了 Equity(通过 RE),但负债端没有对应增加,**L + E 不等式被打破**。 |
|
|
| **修正:** 扩展为 `Accounts_Payable(t) = OpEx(t) + Servicing_Costs(t)`。 |
|
|