lidu16/ethiopia-fi-forecast

# Ethiopia Financial Inclusion - Unified Data Format v2 ## Key Design Principle **Don't force interpretation onto data.** The previous version made a mistake: it assigned events to pillars (e.g., "Telebirr Launch" → USAGE). This is **biased** because: - Telebirr affects both ACCESS and USAGE - Fayda affects ACCESS, GENDER, and TRUST - The pillar assignment is an **interpretation**, not a fact ## The Correct Approach | Record Type | `category` column | `pillar` column | |-------------|-------------------|-----------------| | `observation` | (empty) | **YES** - what dimension is measured | | `target` | (empty) | **YES** - what dimension is the goal | | `event` | **Event type** (policy, product_launch, etc.) | **(empty)** - no pre-assignment | | `impact_link` | (empty) | **YES** - pillar of the affected indicator | --- ## How It Works ### Events are neutral ```csv EVT_0001,,event,product_launch,,Telebirr Launch,... ``` - `category` = what type of event (product_launch) - `pillar` = empty (no pre-interpretation) ### Impact links capture effects ```csv IMP_0001,EVT_0001,impact_link,,ACCESS,...,ACC_OWNERSHIP,direct,increase,high,15,12,... IMP_0003,EVT_0001,impact_link,,USAGE,...,USG_P2P_COUNT,direct,increase,high,25,6,... ``` - One event → multiple impact_links - Each impact_link has a pillar (derived from the affected indicator) ### Query: "What affects ACCESS?" ```python # Get all impact_links that affect ACCESS indicators access_impacts = df[ (df['record_type'] == 'impact_link') & (df['pillar'] == 'ACCESS') ] # Join to get event details access_events = access_impacts.merge( df[df['record_type'] == 'event'], left_on='parent_id', right_on='record_id' ) ``` --- ## Event Categories | category | Description | Examples | |----------|-------------|----------| | `product_launch` | New product/service | Telebirr, M-Pesa | | `market_entry` | New competitor | Safaricom Ethiopia | | `policy` | Government strategy | NFIS-II | | `regulation` | Regulatory directive | K …