LogicGoInfotechSpaces's picture
Update app.py
1b441ae verified
# app.py
import calendar
import math
import os
from collections import defaultdict
from datetime import datetime, timezone
from typing import Dict, List, Optional, Tuple
from time import perf_counter
from bson import ObjectId
from dotenv import load_dotenv
from fastapi import FastAPI, HTTPException
from pydantic import BaseModel, Field
from pymongo import MongoClient
from pymongo.collection import Collection
load_dotenv()
app = FastAPI(title="Expense Prediction API", version="1.0.0")
# ---------- Configurable constants ----------
MAX_HISTORY_MONTHS = int(os.getenv("MAX_HISTORY_MONTHS", "36")) # months to fetch for detection/tuning
SEASONALITY_PERIOD = int(os.getenv("SEASONALITY_PERIOD", "12")) # monthly seasonality (12 months)
SEASONALITY_AMPLITUDE_THRESHOLD = float(os.getenv("SEASONALITY_AMPLITUDE_THRESHOLD", "0.18"))
# grid-search limits (keeps tuning light)
ALPHA_GRID = [0.3, 0.5, 0.7]
BETA_GRID = [0.1, 0.3, 0.5]
GAMMA_GRID = [0.1, 0.3, 0.5]
MAX_GRID_SEARCH_COMBINATIONS = 30 # safety cap
# ------------------------------------------------
class MonthlyExpense(BaseModel):
year: int
month: int
total: float = Field(..., description="Total expenses recorded for the month")
class CategoryPrediction(BaseModel):
headCategoryId: str
title: str
history: List[MonthlyExpense]
predictionMonth: MonthlyExpense
class PredictionResponse(BaseModel):
userId: str
categories: List[CategoryPrediction]
class APIResponse(BaseModel):
status: str
message: str
data: Optional[PredictionResponse] = None
class MongoConnection:
def __init__(self) -> None:
mongo_uri = os.getenv("MONGO_URI")
if not mongo_uri:
raise RuntimeError("MONGO_URI is not configured in the environment")
self._client = MongoClient(mongo_uri, tz_aware=True)
self._database = self._client.get_default_database()
self.transactions: Collection = self._database["transactions"]
self.headcategories: Collection = self._database["headcategories"]
self.api_logs: Collection = self._database["api_logs"]
mongo = MongoConnection()
# ----------------- Date helpers -----------------
def _first_day_of_month(dt: datetime) -> datetime:
return dt.replace(day=1, hour=0, minute=0, second=0, microsecond=0)
def _shift_months(dt: datetime, months: int) -> datetime:
month_index = dt.month - 1 + months
year = dt.year + month_index // 12
month = month_index % 12 + 1
last_day = calendar.monthrange(year, month)[1]
day = min(dt.day, last_day)
return dt.replace(year=year, month=month, day=day)
def month_to_index(year: int, month: int) -> int:
return year * 12 + (month - 1)
def index_to_month(idx: int) -> Tuple[int, int]:
year = idx // 12
month = (idx % 12) + 1
return year, month
def log_api_event(
name: str,
status: str,
response_time: float,
user_id: Optional[str] = None,
error_message: Optional[str] = None,
):
payload = {
"name": name,
"status": status,
"response_time": round(response_time, 3),
"user_id": user_id or "anonymous",
"date": datetime.now(timezone.utc),
}
if error_message:
payload["error_message"] = error_message
try:
mongo.api_logs.insert_one(payload)
except Exception:
# never crash API because of logging
pass
# ------------------------------------------------
# ----------------- Time series utilities -----------------
def build_continuous_series(history: List[MonthlyExpense]) -> Tuple[List[float], List[Tuple[int, int]]]:
"""
Given sparse monthly history items (year, month, total), build a continuous series
covering from earliest to latest month in history. Missing months are represented by None.
Returns (values_list_with_none, list_of_(year,month)_corresponding).
"""
if not history:
return [], []
# sort history
history_sorted = sorted(history, key=lambda h: (h.year, h.month))
start_idx = month_to_index(history_sorted[0].year, history_sorted[0].month)
end_idx = month_to_index(history_sorted[-1].year, history_sorted[-1].month)
length = end_idx - start_idx + 1
idx_to_val = {}
for h in history_sorted:
idx = month_to_index(h.year, h.month)
idx_to_val[idx] = h.total
series = []
months = []
for i in range(start_idx, end_idx + 1):
months.append(index_to_month(i))
series.append(idx_to_val.get(i, None))
return series, months
def impute_missing(series: List[Optional[float]]) -> List[float]:
"""
Fill missing values (None) by linear interpolation. If leading/trailing Nones remain,
forward/backfill with nearest value or 0 if no data.
"""
n = len(series)
if n == 0:
return []
arr = [None if v is None else float(v) for v in series]
# collect indices of non-None
known = [i for i, v in enumerate(arr) if v is not None]
if not known:
# all missing -> return zeros
return [0.0] * n
# linear interpolation between known points
for i in range(len(known) - 1):
a = known[i]
b = known[i + 1]
va = arr[a]
vb = arr[b]
step = (vb - va) / (b - a)
for j in range(a + 1, b):
arr[j] = va + step * (j - a)
# fill leading
first = known[0]
for i in range(0, first):
arr[i] = arr[first]
# fill trailing
last = known[-1]
for i in range(last + 1, n):
arr[i] = arr[last]
return [float(x) for x in arr]
def seasonal_strength(series: List[float], period: int = SEASONALITY_PERIOD) -> float:
"""
Estimate seasonality strength for monthly data.
Returns amplitude_ratio = (max_month_mean - min_month_mean) / overall_mean
Higher value => stronger seasonality.
Requires at least 2 * period data points for a reliable estimate.
"""
n = len(series)
if n < 2 * period:
return 0.0
# compute month-of-year means
month_buckets = [[] for _ in range(period)]
for idx, val in enumerate(series):
month = idx % period
month_buckets[month].append(val)
month_means = [ (sum(b)/len(b)) if b else 0.0 for b in month_buckets ]
overall_mean = sum(series) / len(series) if series else 0.0
if overall_mean == 0:
return 0.0
amplitude = max(month_means) - min(month_means)
return amplitude / overall_mean
# ----------------- Forecasting algorithms -----------------
def holt_double_forecast(series: List[float], alpha: float, beta: float, n_forecast: int = 1) -> List[float]:
"""
Holt's linear method (double exponential smoothing).
Returns list of length n_forecast (forecast ahead).
"""
n = len(series)
if n == 0:
return [0.0] * n_forecast
if n == 1:
return [series[-1]] * n_forecast
level = series[0]
trend = series[1] - series[0]
for t in range(1, n):
value = series[t]
prev_level = level
level = alpha * value + (1 - alpha) * (level + trend)
trend = beta * (level - prev_level) + (1 - beta) * trend
# forecast h steps ahead
forecasts = [level + (i + 1) * trend for i in range(n_forecast)]
return [max(0.0, f) for f in forecasts]
def holt_winters_additive(series: List[float], season_length: int, alpha: float, beta: float, gamma: float, n_forecast: int = 1) -> List[float]:
"""
Additive Holt-Winters seasonal method.
series: list of floats (no missing) where season_length is known (e.g., 12)
"""
n = len(series)
if n == 0:
return [0.0] * n_forecast
if n < season_length * 2:
# not enough data to initialize seasonals reliably -> fallback to holt_double
return holt_double_forecast(series, alpha, beta, n_forecast)
# initialize level, trend, seasonals
seasonals = _initial_seasonal_components(series, season_length)
level = sum(series[:season_length]) / season_length
trend = (sum(series[season_length:2*season_length]) - sum(series[:season_length])) / (season_length * season_length)
result = []
for i in range(n + n_forecast):
if i < n:
val = series[i]
last_level = level
level = alpha * (val - seasonals[i % season_length]) + (1 - alpha) * (level + trend)
trend = beta * (level - last_level) + (1 - beta) * trend
seasonals[i % season_length] = gamma * (val - level) + (1 - gamma) * seasonals[i % season_length]
# in-sample prediction (not used)
else:
# forecast
m = i - n + 1
forecast = level + m * trend + seasonals[i % season_length]
result.append(max(0.0, forecast))
# ensure length matches n_forecast
return result[:n_forecast]
def _initial_seasonal_components(series: List[float], season_length: int) -> List[float]:
"""
Initialize seasonality components by averaging.
"""
seasonals = [0.0] * season_length
n_seasons = len(series) // season_length
if n_seasons == 0:
return seasonals
season_averages = []
for j in range(n_seasons):
start = j * season_length
season_avg = sum(series[start:start + season_length]) / season_length
season_averages.append(season_avg)
for i in range(season_length):
s = 0.0
for j in range(n_seasons):
s += series[j * season_length + i] - season_averages[j]
seasonals[i] = s / n_seasons
return seasonals
# ----------------- Dynamic WMA -----------------
def dynamic_wma(series: List[float], max_len: int = 6) -> float:
"""
Compute a dynamic WMA using up to max_len most recent months.
The weights adapt based on volatility: higher volatility -> smoother (older months get more weight).
"""
n = len(series)
if n == 0:
return 0.0
take = min(n, max_len)
recent = series[-take:]
# compute month-to-month relative changes
if len(recent) >= 2:
changes = [abs(recent[i] - recent[i - 1]) for i in range(1, len(recent))]
vol = sum(changes) / len(changes) if changes else 0.0
else:
vol = 0.0
# base weights favor recent months
base_weights = [ (i + 1) for i in range(take) ] # 1..take
base_weights = list(reversed(base_weights)) # newest highest
total = sum(base_weights)
base_weights = [w/total for w in base_weights]
# adaptation factor: more vol -> flatten weights
# vol_ratio normalized roughly w.r.t average magnitude
avg = sum(recent) / len(recent) if recent else 1.0
vol_ratio = (vol / avg) if avg else 0.0
# clamp vol_ratio
vol_ratio = max(0.0, min(vol_ratio, 1.0))
# blend between base_weights and equal weights
equal_weights = [1.0 / take] * take
blend = min(0.7, vol_ratio) # limit blend to avoid extreme flattening
weights = [(1 - blend) * bw + blend * ew for bw, ew in zip(base_weights, equal_weights)]
# compute prediction
prediction = sum(w * v for w, v in zip(weights, reversed(recent))) # reversed so weights map newest->oldest
return max(0.0, prediction)
# ----------------- Parameter tuning (lightweight) -----------------
def walk_forward_cv_mse(series: List[float], forecast_func, params: dict, min_train_size: int = 6) -> float:
"""
Perform walk-forward validation computing MSE. forecast_func must accept (train_series, params) and return a single-step forecast.
"""
n = len(series)
if n < min_train_size + 1:
# not enough data to validate -> return large error so tuner avoids complex models
return float("inf")
errors = []
# iterate rolling window
for split in range(min_train_size, n):
train = series[:split]
actual = series[split]
try:
pred = forecast_func(train, params)
except Exception:
return float("inf")
if pred is None:
return float("inf")
errors.append((pred - actual) ** 2)
return sum(errors) / len(errors) if errors else float("inf")
def forecast_wrapper_holt(train: List[float], params: dict) -> float:
alpha = params.get("alpha", 0.5)
beta = params.get("beta", 0.3)
return holt_double_forecast(train, alpha, beta, n_forecast=1)[0]
def forecast_wrapper_hw(train: List[float], params: dict) -> float:
alpha = params.get("alpha", 0.5)
beta = params.get("beta", 0.3)
gamma = params.get("gamma", 0.2)
season_length = params.get("season_length", SEASONALITY_PERIOD)
return holt_winters_additive(train, season_length, alpha, beta, gamma, n_forecast=1)[0]
def tune_parameters(series: List[float], seasonal: bool, season_length: int = SEASONALITY_PERIOD) -> dict:
"""
Lightweight grid search for (alpha, beta, gamma) returning best params.
Uses walk-forward CV to score parameter combinations.
"""
best = None
best_score = float("inf")
combos_tested = 0
if seasonal:
grid = []
for a in ALPHA_GRID:
for b in BETA_GRID:
for g in GAMMA_GRID:
grid.append({"alpha": a, "beta": b, "gamma": g, "season_length": season_length})
else:
grid = [{"alpha": a, "beta": b} for a in ALPHA_GRID for b in BETA_GRID]
# cap combos
if len(grid) > MAX_GRID_SEARCH_COMBINATIONS:
grid = grid[:MAX_GRID_SEARCH_COMBINATIONS]
for params in grid:
combos_tested += 1
if seasonal:
score = walk_forward_cv_mse(series, forecast_wrapper_hw, params, min_train_size=max(6, season_length))
else:
score = walk_forward_cv_mse(series, forecast_wrapper_holt, params, min_train_size=6)
if score < best_score:
best_score = score
best = params
if best is None:
# fallback default
if seasonal:
return {"alpha": 0.5, "beta": 0.3, "gamma": 0.2, "season_length": season_length}
else:
return {"alpha": 0.5, "beta": 0.3}
return best
# ----------------- Top-level predictor combining everything -----------------
def _predict_next_month(history: List[MonthlyExpense]) -> float:
"""
Comprehensive predictor:
- builds continuous series and imputes missing months
- auto-detects seasonality
- tunes parameters (lightweight) per series
- uses Holt-Winters if seasonal, else Holt
- fallback to dynamic WMA for very short/noisy series
"""
if not history:
return 0.0
# limit history length to MAX_HISTORY_MONTHS (use most recent months)
history_sorted = sorted(history, key=lambda h: (h.year, h.month))
if len(history_sorted) > MAX_HISTORY_MONTHS:
history_sorted = history_sorted[-MAX_HISTORY_MONTHS:]
# Build continuous series (may contain Nones for missing months)
series_with_none, months = build_continuous_series(history_sorted)
series = impute_missing(series_with_none)
# if after imputation all zeros, return 0
if all(v == 0.0 for v in series):
return 0.0
n = len(series)
# If very short history (<=2) use simple rules / dynamic WMA
if n <= 2:
return round(dynamic_wma(series, max_len=2), 2)
# Seasonality detection: needs at least 2 * season_length samples for reliability
season_strength = seasonal_strength(series, period=SEASONALITY_PERIOD)
is_seasonal = season_strength >= SEASONALITY_AMPLITUDE_THRESHOLD and n >= 2 * SEASONALITY_PERIOD
# If not much data but still some seasonality signal present and we have at least season_length points,
# we can still attempt seasonal HW but with care.
season_length_used = SEASONALITY_PERIOD if is_seasonal else None
# Tuning: per-series personalized coefficients
try:
tuned = tune_parameters(series, seasonal=is_seasonal, season_length=season_length_used or SEASONALITY_PERIOD)
except Exception:
tuned = None
# If tuning failed or not enough data, fallback defaults
if tuned is None:
if is_seasonal:
tuned = {"alpha": 0.5, "beta": 0.3, "gamma": 0.2, "season_length": SEASONALITY_PERIOD}
else:
tuned = {"alpha": 0.5, "beta": 0.3}
# Edge case: if the series is extremely volatile compared to mean, prefer dynamic WMA (more robust)
mean_val = sum(series) / len(series) if series else 0.0
diffs = [abs(series[i] - series[i - 1]) for i in range(1, len(series))] if len(series) >= 2 else [0.0]
avg_diff = sum(diffs) / len(diffs) if diffs else 0.0
volatility_ratio = (avg_diff / mean_val) if mean_val else 0.0
if volatility_ratio > 1.0 and n < 6:
# extremely volatile and short history -> WMA is safer
pred = dynamic_wma(series, max_len=min(6, n))
return round(pred, 2)
# Choose model
if is_seasonal:
alpha = tuned.get("alpha", 0.5)
beta = tuned.get("beta", 0.3)
gamma = tuned.get("gamma", 0.2)
season_length = tuned.get("season_length", SEASONALITY_PERIOD)
pred = holt_winters_additive(series, season_length, alpha, beta, gamma, n_forecast=1)[0]
else:
alpha = tuned.get("alpha", 0.5)
beta = tuned.get("beta", 0.3)
pred = holt_double_forecast(series, alpha, beta, n_forecast=1)[0]
# final safety clamps
if math.isnan(pred) or pred is None or pred < 0:
# fallback to recent avg
pred = sum(series[-3:]) / min(3, len(series))
return round(float(pred), 2)
# ----------------- API endpoint -----------------
@app.get("/users/{user_id}/expense-prediction",response_model=APIResponse,)
def predict_expense(user_id: str):
start_time = perf_counter()
try:
user_object_id = ObjectId(user_id)
except Exception:
log_api_event(
name="Expense Prediction",
status="failed",
response_time=0,
user_id=user_id,
error_message="Invalid user id",
)
raise HTTPException(status_code=400, detail="Invalid user id")
try:
now = datetime.now(timezone.utc)
start_period = _shift_months(_first_day_of_month(now), -MAX_HISTORY_MONTHS + 1)
prediction_month = _shift_months(_first_day_of_month(now), 1)
pipeline = [
{
"$match": {
"user": user_object_id,
"type": "EXPENSE",
"headCategory": {"$ne": None},
"date": {"$gte": start_period},
}
},
{
"$project": {
"amount": 1,
"headCategory": 1,
"year": {"$year": "$date"},
"month": {"$month": "$date"},
}
},
{
"$group": {
"_id": {
"headCategory": "$headCategory",
"year": "$year",
"month": "$month",
},
"total": {"$sum": "$amount"},
}
},
{
"$lookup": {
"from": "headcategories",
"localField": "_id.headCategory",
"foreignField": "_id",
"as": "headCategoryDoc",
}
},
{"$unwind": "$headCategoryDoc"},
{"$sort": {"_id.headCategory": 1, "_id.year": 1, "_id.month": 1}},
]
results = list(mongo.transactions.aggregate(pipeline))
grouped: Dict[ObjectId, Dict[str, List[MonthlyExpense]]] = defaultdict(lambda: {"history": []})
for item in results:
head_category_id: ObjectId = item["_id"]["headCategory"]
category_record = grouped[head_category_id]
category_record["title"] = item["headCategoryDoc"].get("title", "Unknown")
category_record["history"].append(
MonthlyExpense(
year=item["_id"]["year"],
month=item["_id"]["month"],
total=float(item["total"]),
)
)
categories: List[CategoryPrediction] = []
for head_category_id, record in grouped.items():
history = sorted(record["history"], key=lambda doc: (doc.year, doc.month))
predicted_total = _predict_next_month(history)
categories.append(
CategoryPrediction(
headCategoryId=str(head_category_id),
title=record.get("title", "Unknown"),
history=history,
predictionMonth=MonthlyExpense(
year=prediction_month.year,
month=prediction_month.month,
total=predicted_total,
),
)
)
response_data = PredictionResponse(userId=user_id, categories=categories)
log_api_event(
name="Expense Prediction",
status="success",
response_time=perf_counter() - start_time,
user_id=user_id,
)
return APIResponse(
status="success",
message="Expense prediction generated successfully",
data=response_data,
)
except Exception as exc:
log_api_event(
name="Expense Prediction",
status="failed",
response_time=perf_counter() - start_time,
user_id=user_id,
error_message=str(exc),
)
raise HTTPException(status_code=500, detail="Internal server error")
@app.get("/health")
def health():
try:
mongo._client.admin.command("ping")
return {
"status": "ok",
"message": "Service is healthy",
"timestamp": datetime.now(timezone.utc),
}
except Exception as exc:
raise HTTPException(
status_code=503,
detail={
"status": "down",
"message": "Database connectivity failed",
"error": str(exc),
},
)
# import calendar
# import os
# from collections import defaultdict
# from datetime import datetime, timezone
# from typing import Dict, List
# from bson import ObjectId
# from dotenv import load_dotenv
# from fastapi import FastAPI, HTTPException
# from pydantic import BaseModel, Field
# from pymongo import MongoClient
# from pymongo.collection import Collection
# load_dotenv()
# app = FastAPI(title="Expense Prediction API", version="1.0.0")
# class MonthlyExpense(BaseModel):
# year: int
# month: int
# total: float = Field(..., description="Total expenses recorded for the month")
# class CategoryPrediction(BaseModel):
# headCategoryId: str
# title: str
# history: List[MonthlyExpense]
# predictionMonth: MonthlyExpense
# class PredictionResponse(BaseModel):
# userId: str
# categories: List[CategoryPrediction]
# class MongoConnection:
# def __init__(self) -> None:
# mongo_uri = os.getenv("MONGO_URI")
# if not mongo_uri:
# raise RuntimeError("MONGO_URI is not configured in the environment")
# self._client = MongoClient(mongo_uri, tz_aware=True)
# self._database = self._client.get_default_database()
# self.transactions: Collection = self._database["transactions"]
# self.headcategories: Collection = self._database["headcategories"]
# mongo = MongoConnection()
# def _first_day_of_month(dt: datetime) -> datetime:
# return dt.replace(day=1, hour=0, minute=0, second=0, microsecond=0)
# def _shift_months(dt: datetime, months: int) -> datetime:
# month_index = dt.month - 1 + months
# year = dt.year + month_index // 12
# month = month_index % 12 + 1
# last_day = calendar.monthrange(year, month)[1]
# day = min(dt.day, last_day)
# return dt.replace(year=year, month=month, day=day)
# # -----------------------------------------------------------
# # NEW: Weighted Moving Average-based prediction function
# # -----------------------------------------------------------
# def _predict_next_month(history: List[MonthlyExpense]) -> float:
# """Predict next month's expense using Weighted Moving Average (WMA)."""
# totals = [h.total for h in history]
# # Only one month β†’ Just repeat last month
# if len(totals) == 1:
# return round(totals[-1], 2)
# # Two months β†’ Slight smoothing
# if len(totals) == 2:
# last, prev = totals[-1], totals[-2]
# prediction = last * 0.7 + prev * 0.3
# return round(prediction, 2)
# # Three or more months β†’ Use 3-month WMA (0.5, 0.3, 0.2)
# last3 = totals[-3:]
# weights = [0.2, 0.3, 0.5] # oldest β†’ newest
# prediction = sum(v * w for v, w in zip(last3, weights))
# return round(prediction, 2)
# # -----------------------------------------------------------
# # EXPENSE PREDICTION ENDPOINT
# # -----------------------------------------------------------
# @app.get("/users/{user_id}/expense-prediction", response_model=PredictionResponse)
# def predict_expense(user_id: str) -> PredictionResponse:
# try:
# user_object_id = ObjectId(user_id)
# except Exception as exc:
# raise HTTPException(status_code=400, detail="Invalid user id") from exc
# now = datetime.now(timezone.utc)
# start_period = _shift_months(_first_day_of_month(now), -2)
# prediction_month = _shift_months(_first_day_of_month(now), 1)
# pipeline = [
# {
# "$match": {
# "user": user_object_id,
# "type": "EXPENSE",
# "headCategory": {"$ne": None},
# "date": {"$gte": start_period},
# }
# },
# {
# "$project": {
# "amount": 1,
# "headCategory": 1,
# "year": {"$year": "$date"},
# "month": {"$month": "$date"},
# }
# },
# {
# "$group": {
# "_id": {
# "headCategory": "$headCategory",
# "year": "$year",
# "month": "$month",
# },
# "total": {"$sum": "$amount"},
# }
# },
# {
# "$lookup": {
# "from": "headcategories",
# "localField": "_id.headCategory",
# "foreignField": "_id",
# "as": "headCategoryDoc",
# }
# },
# {"$unwind": "$headCategoryDoc"},
# {"$sort": {"_id.headCategory": 1, "_id.year": 1, "_id.month": 1}},
# ]
# results = list(mongo.transactions.aggregate(pipeline))
# grouped: Dict[ObjectId, Dict[str, List[MonthlyExpense]]] = defaultdict(
# lambda: {"history": []}
# )
# for item in results:
# head_category_id: ObjectId = item["_id"]["headCategory"]
# category_record = grouped[head_category_id]
# category_record["title"] = item["headCategoryDoc"].get("title", "Unknown")
# category_record["history"].append(
# MonthlyExpense(
# year=item["_id"]["year"],
# month=item["_id"]["month"],
# total=float(item["total"]),
# )
# )
# categories: List[CategoryPrediction] = []
# for head_category_id, record in grouped.items():
# history = sorted(record["history"], key=lambda doc: (doc.year, doc.month))
# predicted_total = _predict_next_month(history)
# categories.append(
# CategoryPrediction(
# headCategoryId=str(head_category_id),
# title=record.get("title", "Unknown"),
# history=history,
# predictionMonth=MonthlyExpense(
# year=prediction_month.year,
# month=prediction_month.month,
# total=predicted_total,
# ),
# )
# )
# return PredictionResponse(userId=user_id, categories=categories)